CN100367347C - Speech encoder and speech decoder - Google Patents

Abstract

An excitation vector generation apparatus comprising: a pulse vector generation unit having N channels (N ≧ 1) for generating pulse vectors; corresponding to N channels, each channel stores M (M is more than or equal to 1) storage units with diffusion modes; a selection unit selectively taking out the diffusion pattern from the memory unit for each channel; each channel carries out convolution operation of the extracted diffusion mode and the generated pulse vector and generates N diffusion units of diffusion vectors; and an excitation vector generation unit for generating excitation vectors from the generated N diffusion vectors.

Description

Voice signal encoder and voice signal decoder

technical field

The present invention relates to a speech signal encoder and a speech signal decoder for efficiently encoding and decoding speech information.

Background technique

Today, speech coding techniques for efficiently coding and decoding speech information are being developed. "Code Excited Linear Prediction: High Quality Speech at Low Bit Rate" (Code Excited Linear Prediction: High Quality Speech at Low Bit Rate) (written by M.R.Schroeder; published in ICASSP'85, pp.937～940) records that based on this speech Coding technology CELP type speech signal coder. This voice signal encoder performs linear prediction on each frame obtained by dividing the input voice with a fixed time, and obtains the prediction residual (excitation signal) from the linear prediction of each frame, and stores the adaptive codebook of the driving sound source in the past and stores multiple A random codebook of random code vectors encodes the prediction residual.

FIG. 1 shows a functional block diagram of a conventional CELP type speech signal encoder (hereinafter simply referred to as "speech encoder").

The linear predictive analysis unit 12 performs linear predictive analysis on the speech signal 11 input to the CELP type speech signal encoder. Using this linear predictive analysis, a linear predictive coefficient can be obtained. The linear prediction coefficient is a parameter representing the spectrum envelope characteristic of the speech signal 11 . The linear prediction coefficient obtained by the linear prediction analysis unit 12 is quantized in the linear prediction coefficient coding unit 13 , and the quantized linear prediction coefficient is sent to the linear prediction coefficient decoding unit 14 . The quantization number obtained by quantization is also input to the encoding output unit 24 as a linear predictive code. The linear prediction coefficient decoding unit 14 decodes the quantized linear prediction coefficient obtained by the linear prediction coefficient encoding unit 13 to obtain the coefficient of the synthesis filter, and the linear prediction coefficient decoding unit 14 outputs the coefficient of the synthesis filter to the synthesis filter 15 .

The adaptive codebook 17 is a codebook for outputting multiple kinds of candidate adaptive codevectors, and is composed of a buffer for storing past multi-frame driving sound sources. The adaptive code vector is a time series vector representing the periodic component in the input speech.

The random codebook 18 is a codebook storing multiple kinds of candidate random code vectors, the type of which corresponds to the number of allocated bits. The random code vector is a time-series vector representing the aperiodic component of the input speech.

After the adaptive code gain weighting unit 19 and the random code gain weighting unit 20 multiply the candidate vectors output by the adaptive codebook 17 and the random codebook 18 by the adaptive gain and the random code gain read from the weighted codebook 21, output to adder 22.

The weighted codebook is a kind of memory, which respectively stores various weights to be multiplied with candidate adaptive code vectors and weighted numbers to be multiplied with candidate random code vectors, the types of which correspond to the number of allocated bits.

The adder 22 adds the candidate adaptive code vector and the candidate random code vector weighted by the adaptive code gain weighting unit 19 and the random code gain weighting unit 20 respectively to generate a candidate driving excitation vector, and outputs it to the synthesis filter 15 .

The synthesis filter 15 is an omnipolar filter composed of synthesis filter coefficients obtained by the linear prediction coefficient decoding unit 14 . The synthesis filter 15 has a function of outputting a candidate synthesized voice vector when a candidate driving sound source vector from the adder 22 is input.

The distortion calculation unit 16 calculates the distortion between the output of the synthesis filter 15 (namely, the candidate synthesized speech vector) and the input speech 11, and outputs the obtained distortion value to the code number specifying unit 23. The code number specifying unit 23 specifies three kinds of code numbers (adaptive code number, random code number and and weight code number). Then, the three code numbers specified by the code number specifying unit 23 are output to the code output unit 24 . The coding output unit 24 combines the linear prediction code number obtained by the linear prediction coefficient coding unit 13 and the adaptive code number, random code number and weighted code number specified by the code number specification unit 23, and outputs to the transmission line.

FIG. 2 shows a functional block diagram of a CELP type speech signal decoder (hereinafter simply referred to as "speech decoder") for decoding a signal encoded by the above encoder. In this voice signal decoder, code input unit 31 receives the code that voice signal coder (Fig. 1) sends, and the code that receives is decomposed into linear predictive code number, adaptive code number, noise code number and weighted code number, And the decomposed codes are respectively output to the linear prediction coefficient decoding unit 32 , the adaptive codebook 33 , the noise codebook 34 and the weighted codebook 35 .

Next, the linear predictive coefficient decoding unit 32 decodes the linear predictive code number obtained by the encoding input unit 31 to obtain a synthesis filter coefficient, and outputs it to the synthesis filter 39 . Then, read the adaptive code vector from the position corresponding to the adaptive code number in the adaptive codebook, read the random code vector corresponding to the random code number from the random codebook, and then read the weighted code number corresponding to Adaptive code gain and random code gain of . Furthermore, after the adaptive code weighting unit 36 multiplies the adaptive code vector by the adaptive code gain, it is sent to the adder 38 . Similarly, after the random code vector weighting unit 37 multiplies the random code vector by the random code gain, it is sent to the adder 38 .

Adder 38 generates the driving sound source vector after the above-mentioned two coding vectors are added, and the driving sound source that produces is sent to adaptive codebook 33, to update the register, and this driving sound source is also sent to synthesis filter 39 to drive filter. The synthesis filter 39 is driven by the driving sound source vector obtained by the adder 38, and uses the output of the linear prediction coefficient decoding unit 32 to reproduce the synthesized speech.

The distortion calculation unit 16 of the CELP type voice signal encoder generally utilizes the following formula (formula 1) to calculate the required distortion E:

E=‖V-(gaHP+gcHC)‖ ² (1)

V: input voice signal (vector)

H: synthesis filter impulse response convolution matrix

Among them, h is the impulse response (vector) of the synthesis filter, L is the frame length,

p: adaptive code vector

c: random code vector

ga: adaptive code gain

gc: noise code gain

Here, in order to minimize the distortion E of the expression (1), it is necessary to calculate the distortion in a closed loop for all combinations of adaptive code numbers, random code numbers, and weighted code numbers, and specify each code number.

However, if the closed-loop retrieval of formula (1) is performed, the amount of calculation processing is too large. Therefore, generally, the adaptive code number is firstly specified by vector quantization using the adaptive codebook, and secondly, the random code number is specified by vector quantization using the random codebook. Finally, The weighted code numbers are specified by vector quantization using a weighted codebook. In this case, the vector quantization process using the random codebook will be further described in detail.

When the adaptive code number and the adaptive code gain are determined in advance or temporarily, the distortion estimating formula of the formula (1) becomes the following formula (2).

Ec=‖X-gcHC‖ (2)

Wherein, the vector X in the formula (2) adopts the adaptive code number and adaptive code gain specified in advance or provisionally, and the noise source information obtained by the following formula (3) (the target vector used for specifying the noise code number ).

X＝V-gaHP (3)

ga: adaptive code gain

V: voice signal (vector)

H: synthesis filter impulse response convolution matrix

P: adaptive code vector

In the case of specifying the random code gain gc after specifying the random code number, it is generally known that gc in formula (2) can take any value, so that the processing of formula (2) to minimize the specified random code vector number (the vector of noise source information Quantization processing) can be replaced by a random code vector number that maximizes the fractional expression of the following equation (4).

$\frac{{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}^{<span\; class="notranslate">\; \&prime;</span>}\mathrm{<span\; class="notranslate">\; Hc</span>}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; Hc</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; )</span>$

That is, when the adaptive code number and the adaptive code gain have previously been or are temporarily specified, the noise source information vector quantization process is to specify the candidate noise code vector number that maximizes the fractional expression of Equation (4) calculated by the distortion calculation unit 16 processing.

In the early CELP type encoder/decoder, data obtained by storing random number sequences corresponding to the number of allocated bits in the memory is used as a random codebook. However, there are problems in that a very large storage capacity is required, and at the same time, the calculation processing amount for calculating the distortion of Equation (4) for each random code vector candidate is enormous.

As a method for solving this problem, "8kb/s ACELP speech coding using 10ms speech frame: Candidate CCITT standard" ("8KBIT/S ACELP CODING OF SPEECH WITH 10MS SPEECH-FRAME: A CANDIDATE FOR CCITT STANDARDIZATION ") (by R.Salami.C.Laflamme and J-P.Adoul, published in ICASSP'94, pp.II-97～II-100, 1994), etc., using an algebraic sound source that generates a sound source vector algebraically CELP type speech signal encoder/decoder for vector generation unit.

Yet, in the CELP type voice signal encoder/decoder that adopts above-mentioned algebraic sound source generating unit in random code book, often with a small amount of pulse approximation expression by the noise source information that formula (3) obtains (specify the target vector of random code number), Therefore, there is a limit in seeking to improve voice quality. When actually looking at the elements of the noise source information X in equation (3), it is almost never the case that these elements are composed of only a small number of pulses. This shows that there are limitations.

Contents of the invention

The purpose of the present invention is to provide a new sound source vector generating device, which can generate sound source vectors with high shape and statistical similarity to the sound source vectors obtained when actually analyzing speech signals.

Another object of the present invention is to provide a kind of CELP speech signal coder/decoder, speech signal communication system and speech signal recording system, they can obtain than algebraic sound source generation unit by using above-mentioned source of sound vector generation device as noise code book Synthetic speech of high quality when used as a noise codebook.

The first aspect of the present invention is a sound source vector generation device, which is characterized in that it includes: pulse vector generation with N (N≥1) channels for generating a pulse vector with a polar unit pulse on a certain element on the vector axis Unit; has the function of storing M (M ≥ 1) kinds of diffusion patterns for each channel in the N, and has a diffusion pattern storage and selection unit capable of selecting a diffusion pattern from the stored M diffusion patterns; The pulse vector outputted by the pulse vector generation unit and the convolution operation of the diffusion mode selected by the diffusion mode storage selection unit, and generate N diffusion vectors; a pulse vector diffusion unit with the function of generating N diffusion vectors; After the N diffusion vectors are added, the function of generating the sound source vector is a diffusion vector adder. Make described impulse vector generating unit have the function of producing N impulse vectors (N≥1) in algebraic mode, add that described diffusion mode storage selection unit pre-stores the diffusion obtained by learning the shape (characteristic) of actual voice vector in advance mode, so it can generate a sound source vector whose shape is closer to the shape of the actual sound source vector than the previous algebraic sound source generation unit.

A second aspect of the present invention is a CELP speech signal encoder/decoder characterized in that said sound source vector generator is used in a random codebook. Compared with the speech signal encoder/decoder using the algebraic sound source generation unit in the conventional noise codebook, it can generate the sound source vector closer to the actual shape, so the speech signal encoder/decoder that can output the synthesized speech with higher quality can be obtained. Decoder, voice signal communication system and voice signal recording system.

Description of drawings

Fig. 1 is a functional block diagram of a conventional CELP type speech signal encoder.

Fig. 2 is a functional block diagram of a conventional CELP type voice signal decoder.

Fig. 3 is a functional block diagram of the sound source vector generating device according to the first embodiment of the present invention.

Fig. 4 is a functional block diagram of a CELP type speech signal encoder according to a second embodiment of the present invention.

Fig. 5 is a functional block diagram of a CELP type voice signal decoder according to the second embodiment of the present invention.

Fig. 6 is a functional block diagram of a CELP type speech signal encoder according to a third embodiment of the present invention.

Fig. 7 is a functional block diagram of a CELP type speech signal encoder according to a fourth embodiment of the present invention.

Fig. 8 is a functional block diagram of a CELP type speech signal encoder according to a fifth embodiment of the present invention.

Fig. 9 is a block diagram of the vector quantization function in the fifth embodiment.

Fig. 10 is an explanatory diagram of an algorithm for extracting objects in the fifth embodiment.

Fig. 11 is a functional block diagram of predictive quantization in the fifth embodiment.

Fig. 12 is a functional block diagram of predictive quantization in the sixth embodiment.

Fig. 13 is a functional block diagram of a CELP type speech signal encoder in the seventh embodiment.

Fig. 14 is a functional block diagram of a distortion calculation unit in the seventh embodiment.

Detailed ways

Embodiments of the present invention will be described below using the drawings.

(first embodiment)

Fig. 3 is a functional block diagram of the sound source vector generation device according to the embodiment of the present invention. This sound source vector generating device includes: a pulse vector generating unit 101 with multiple channels; a diffusion mode storage selection unit 102 with a diffusion mode storage unit and a switch; a pulse vector diffusion unit 103 of a diffusion pulse vector; a plurality of channel pulses to be diffused Diffusion vector adder 104 for vector addition.

The pulse vector generation section 101 has N channels (in the present embodiment, the case of N=3 will be described), and these channels generate a vector (hereinafter referred to as a pulse vector) in which a unit pulse with a polarity is arranged in a certain element on the vector axis.

The diffusion pattern storage selection unit 102 has memory cells M1-M3 and switches SW1-SW3, the former stores M kinds of diffusion patterns for each channel (the case of M=2 will be described in this embodiment), and the latter selects from each memory cell M1 ~M3 respectively select one of the M diffusion modes.

The pulse vector diffusion unit 103 performs convolution operation of the pulse vector output by the pulse vector generation unit 101 and the diffusion pattern output by the diffusion pattern storage selection unit 102 for each channel, and generates N diffusion vectors.

The diffuse vector adder 104 adds the N diffuse vectors generated by the pulse vector diffuser 103 to generate the sound source vector 105 .

In this embodiment, the case where the pulse vector generation unit 101 algebraically generates N pulse vectors (N=3) according to the rules described in Table 1 below will be described.

Table 1

The operation of the sound source vector generator configured as described above will be described. The diffusion pattern storage selection section 102 selects one of the two diffusion patterns stored for each channel, and outputs it to the pulse vector diffusion section 103 . However, numbers are assigned exclusively corresponding to the selected diffusion pattern combinations (the total number of combinations M ^N =8 types).

Next, pulse vector generating section 101 algebraically generates pulse vectors corresponding to the number of channels (3 in this embodiment) according to the rules described in Table 1.

The pulse vector diffusion unit 103 performs convolution operation on the diffusion pattern selected by the diffusion pattern storage selection unit 102 and the pulse generated by the pulse vector generation unit 101 using equation (5) to generate a diffusion vector for each channel.

$\mathrm{<span\; class="notranslate">\; ci</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; L</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; wij</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; di</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; )</span>$

Among them, n: 0～L-1

L: Diffusion vector length

i: channel number

j: Diffusion pattern number (j=1~M)

ci: Diffusion vector for channel i

wij: the jth diffusion mode of channel i

The vector length of Wij(m) is 2L-1(m: -(L-1)~L-1), but among the 2L-1 elements, the Lij element is the element that can specify the value, and the other elements are zero

di: pulse vector of channel i

di=±δ(n-pi), n=0～L-1

pi: alternate pulse vector of channel i

The diffusion vector adder 104 adds the three diffusion vectors generated by the pulse vector diffusion unit 103 using the formula (6), and generates the sound source vector 105 .

$\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; ci</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 6</span>}<span\; class="notranslate">\; )</span>$

c: sound source vector

ci: diffusion vector

i: channel number (i=1～N)

n: vector element number (n=0～L-1, where L is the length of the sound source vector)

The sound source vector generation device configured in this way can generate various sound source vectors by changing the combination method of the diffusion pattern selected by the diffusion pattern storage selection unit 102 and the pulse position and polarity in the pulse vector generated by the pulse vector generation unit 101. .

Then, the sound source vector generation device configured as above stores two types of diffusion patterns, such as a combination method of the diffusion pattern selected by the selection section 102 and a combination method of the pulse vector shape (pulse position and pulse polarity) generated by the pulse vector generation section 101. Information can be pre-assigned one-to-one corresponding numbers. The diffusion pattern storage selection unit 102 may also perform pre-learning according to actual sound source information, and pre-store the diffusion pattern obtained from the learning result.

If adopt above-mentioned source of sound vector generation device in the source of sound information generation unit of speech signal coder/decoder, then store the combination number of the combination number of the selected diffusion mode of the selection unit and the combination number of the pulse vector generated by the pulse vector generation unit by transmitting the diffusion mode (can Two types of numbers such as pulse position and pulse polarity are specified. The transmission of noise source information can be realized.

Also, when using the above-configured sound source vector generating means, it is possible to generate an sound source vector whose shape (characteristic) is similar to the actual sound source information compared to using an algebraically generated pulse sound source.

In this embodiment, the case where the diffusion pattern storage and selection section 102 stores two diffusion patterns for each channel is described, but the same action and effect can be obtained when other than two diffusion patterns are allocated to each channel.

In this embodiment, the case where the pulse vector generation unit 101 consists of three channels and is based on the pulse generation rules listed in Table 1 will be described. However, when the number of channels is different, or when the pulse generation rules other than those listed in Table 1 are used , can also achieve the same function and effect.

Furthermore, by constructing a speech signal communication system or a speech signal recording system having the above-mentioned sound source vector generating device or speech signal encoder/decoder, the functions and effects possessed by the above-mentioned sound source vector generating device can be obtained.

(Second Embodiment)

FIG. 4 shows a functional block diagram of a CELP-type voice signal encoder of this embodiment, and FIG. 5 shows a functional block diagram of a CELP-type voice signal decoder of this embodiment.

In the CELP-type speech signal encoder of this embodiment, the excitation vector generator described in the first embodiment is applied to the noise codebook of the CELP-type speech signal encoder in FIG. 1 above. In the CELP type speech signal decoder related to this embodiment, the sound source vector generator of the first embodiment is applied to the noise codebook of the CELP speech signal decoder in FIG. 2 above. Therefore, the processing other than the noise source information vector quantization processing is the same as that of the above-mentioned devices in FIG. 1 and FIG. 2 . In this embodiment, the speech signal encoder and the speech signal decoder will be described centering on the noise source information vector quantization process. Also, as in the first embodiment, the number of channels N=3, the number of diffusion modes per channel M=2, and the generation of pulse vectors is based on FIG. 1 .

The noise source vector quantization processing in the voice signal encoder of Fig. 4 is a process of specifying two types of numbers (diffusion pattern combination number, pulse position and pulse polarity combination number) that maximize the reference value of formula (4).

When the sound source vector generating device in Fig. 3 is used as a noise codebook, the combination numbers of diffusion patterns (8 types) and pulse vector combinations (16384 types when polarity is considered) are specified in a closed loop.

Therefore, the diffusion pattern storage selection section 215 first selects one of the two diffusion patterns stored in itself, and outputs it to the pulse vector diffusion section 217 . Then, pulse vector generating section 216 algebraically generates pulse vectors corresponding to the number of channels (three in this embodiment) according to the rule in FIG. 1 , and outputs the pulse vectors to pulse vector spreading section 217 .

The pulse vector diffusion unit 217 uses the convolution operation of formula (5) to generate a diffusion vector for each channel by using the diffusion mode selected by the diffusion mode storage selection unit 215 and the pulse vector generated by the pulse vector generation unit 216 .

The diffuse vector adder 218 adds the diffuse vectors obtained by the pulse vector spreading unit 217 to generate an excitation vector (which becomes a candidate random code vector).

Then, distortion calculating section 206 calculates the value of Equation (4) of the candidate random code vector obtained by using diffuse vector adder 218 . For all the combinations of the pulse vectors produced by the rules of Table 1, the calculation of the value of the above formula (4) is carried out, and the diffusion mode combination number, the pulse vector combination number (pulse position and combination of its polarity), and the maximum value at that time is output to the code number specifying unit 213.

Next, the diffusion pattern storage selection unit 215 selects a diffusion pattern different from the combination just previously selected from the obtained stored diffusion patterns. Then, for the newly selected and changed combination of diffusion patterns, the value of Equation (4) is calculated for all pulse vector combinations generated by pulse vector generating section 216 according to the rules in Table 1, as described above. The diffusion pattern combination number, the pulse vector combination number, and the maximum value when the formula (4) is maximum are output to the code number specifying unit 213 again.

The above processing is repeated for all the combinations that can be selected from the diffusion patterns stored in the diffusion pattern storage selection unit 215 (the total number of combinations is 8 in the description of this embodiment).

The code number specifying unit 213 compares all 8 vector maximum values calculated by the distortion calculation unit 206, selects the largest one, specifies two combination numbers (diffusion mode combination number, pulse vector combination number) when the maximum value is generated, and uses it as noise The code number is output to the code output unit 214 .

On the other hand, in the speech signal decoder of Fig. 5, the code input unit 301 receives the code sent by the speech signal coder (Fig. 4), and decomposes the received code into corresponding linear prediction number, adaptive code number, noise code number (consisting of two kinds of diffusion mode combination number and pulse vector combination number) and weighted code number, and the code obtained by decomposing is output to linear prediction coefficient decoding unit 302, adaptive codebook 303, noise codebook 304 and Weighted codebook 305 .

Among the random code numbers, the combination number of the diffusion mode is output to the storage selection unit 311 of the diffusion mode, and the combination number of the pulse vector is output to the generation unit 312 of the pulse vector.

Then, linear prediction coefficient decoding section 302 decodes the linear prediction code number to obtain synthesis filter coefficients, and outputs them to synthesis filter 309 . In adaptive codebook 303, adaptive code vectors are read from positions corresponding to adaptive code numbers.

In the noise codebook 304, the diffusion mode storage selection unit 311 reads out the diffusion mode corresponding to the diffusion pulse combination number for each channel and outputs it to the pulse vector diffusion unit 313; the pulse vector generation unit 312 generates the combination of the channel number and the pulse vector No. corresponding pulse vector and output to pulse vector diffusion unit 313; Pulse vector diffusion unit 313 will receive the diffusion pattern from diffusion pattern storage selection unit 311 and the pulse vector received from pulse vector generation unit 312 with formula (5) The convolution operation produces a diffuse vector, which is output to the diffuse vector adder 314 . The spread vector adder 314 adds the spread vectors of each channel generated by the pulse vector spread unit 313 to generate a random code vector.

Read out the adaptive code gain and the random code gain corresponding to the weighted code number from the weighted code book 305, and multiply the adaptive code vector by the adaptive code gain at the adaptive code vector weighting unit 306, and similarly, at the random code weighting unit 307 multiplies the random code vector by the random code gain, and sends it to the adder 308 .

The adder 308 adds the above-mentioned 2 code vectors that have been multiplied by the gain to generate a driving source vector, and outputs the generated driving source vector to the adaptive codebook 303 so as to update the register, and also outputs to the synthesis filter 309, in order to drive the synthesis filter.

The synthesis filter 309 is driven by the driving sound source vector obtained by the adder 308 to reproduce the synthesized speech 310 . Also, the adaptive codebook 303 updates the buffer with the driving excitation vector received from the adder 308 .

But, the diffusion mode storage selection unit in Fig. 4 and Fig. 5 is taken as the distortion calculation reference formula of formula (7) obtained by substituting the sound source vector recorded in formula (6) in C in formula (2) as a cost function, and After learning in advance to make the value of the cost function smaller, the learned diffusion pattern is stored for each channel.

Through the above operations, the sound source vector whose shape is similar to the shape of the actual noise source information (vector X in formula (4)) can be generated, so it is similar to the CELP speech signal encoder/decoder using the algebraic sound source vector generation unit in the noise codebook than, high-quality synthetic speech can be obtained.

$\mathrm{<span\; class="notranslate">\; Ec</span>}<span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; wxya</span>}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; Ci</span>}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}$

$<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}^{\mathrm{<span\; class="notranslate">\; L</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; wxya</span>}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; N</span>}}\mathrm{<span\; class="notranslate">\; Ci</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span>{<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}$

$<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}^{\mathrm{<span\; class="notranslate">\; L</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; wxya</span>}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}^{\mathrm{<span\; class="notranslate">\; L</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}\mathrm{<span\; class="notranslate">\; Wij</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; di</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 7</span>}<span\; class="notranslate">\; )</span>$

X: Specifies the target vector for the noise code number

gc: noise code gain

H: synthesis filter impulse response convolution matrix

C: random code vector

i: channel number (i=1～N)

j: Diffusion pattern number (i=1~M)

ci: Diffusion vector for channel i

Wij: the jth diffusion mode of channel i

di: pulse vector of channel i

L: sound source vector length (n=0～L-1)

In this embodiment, each channel pre-stores M diffusion patterns obtained by pre-learning to make the cost function value of formula (7) smaller. If each channel is made to store at least one diffusion pattern obtained through learning in advance, the function and effect of improving the quality of synthesized voice can also be obtained in this case.

The situation described in this embodiment is that according to all the combinations of the diffusion patterns stored by the diffusion pattern storage selection unit and all the combinations of the candidate pulse vector positions generated by the pulse vector generation unit 6, the reference value in the formula (4) is maximized with a closed-loop regulation combination number. However, the same action and effect can also be obtained by performing preselection based on parameters obtained before specifying the random codebook number (ideal gain of adaptive code vector, etc.), or performing search by open loop.

Furthermore, by configuring a speech signal communication system or a speech signal recording system having the above-mentioned speech signal encoder/decoder, the functions and effects of the sound source vector generation device described in the first embodiment can be obtained.

(third embodiment)

Fig. 6 shows a functional block diagram of a CELP type speech signal encoder according to this embodiment. In this embodiment, in the CELP speech coder adopting the sound source vector generating device of the first embodiment in the noise codebook, the ideal adaptive code gain value obtained before searching the noise codebook is used to perform the diffusion pattern stored in the selection unit. preselection. Except for the peripheral part of the random codebook, it is the same as the CELP type speech signal encoder in Fig. 4. Therefore, this embodiment describes the vector quantization processing of noise source information in the CELP type speech signal encoder of FIG. 6.

This CELP type voice signal encoder has an adaptive codebook 407, an adaptive code gain weighting unit 409, a random codebook 408 composed of the sound source vector generation device described in Embodiment 1, a random code gain weighting unit 410, and a synthesis filter 405. , a distortion calculation unit 406 , a code number specification unit 413 , a diffusion pattern storage selection unit 415 , an impulse vector generation unit 416 , an impulse vector diffusion unit 417 , a diffusion vector adder 418 and an adaptive gain determination unit 419 .

However, in this embodiment, at least one of the M types of diffusion patterns (M≥2) stored in the diffusion pattern storage selection unit 415 is learned in advance, so that the quantization distortion generated when the noise source information is vectorized is small, and by this Diffusion patterns obtained as a result of learning.

For ease of description, the channel number N of the pulse vector generation unit is set to be 3 in the present embodiment, and the diffusion pulse type number M of each channel stored in the diffusion mode storage selection unit is 2, and the following situations are taken for illustration: M kinds of diffusion modes One of (M=2) is the diffusion pattern obtained by the above-mentioned learning, and the other is a random vector string (hereinafter referred to as a random pattern) generated by the random number vector generator. Incidentally, the above-mentioned diffusion pattern obtained by learning is clearly a pulse-like diffusion pattern with a relatively short length, like W11 in FIG. 3 .

In the CELP type speech signal encoder of FIG. 6, processing for specifying an adaptive codebook number is performed before vector quantization of noise source information. Therefore, the vector number (adaptive code number) of the adaptive codebook and the ideal adaptive code gain (temporarily determined) can be referred to when the noise source information vector quantization process is performed. In this embodiment, the ideal adaptive code gain value among them is used for the preselection of the diffused pulses.

Specifically, first, immediately after the adaptive codebook search is completed, the ideal value of the adaptive code gain held by the code number specifying section 413 is output to the distortion calculating section 406 . Distortion calculating section 406 outputs the adaptive code gain received from code number specifying section 413 to adaptive gain determining section 419 .

The adaptive gain determination unit 419 compares the ideal adaptive gain value received from the distortion calculation unit 409 with a preset threshold value. Next, adaptive gain determining section 419 sends a preliminarily used control signal to diffusion pattern storage selecting section 415 based on the result of the magnitude comparison. The content of the control signal indicates that when the gain of the adaptive code is large in the above size comparison, it indicates to select the diffusion mode obtained by performing pre-learning and making the quantization distortion generated during the vector quantization of the noise source information smaller, and the adaptive code in the above size comparison When the gain is not large, it indicates that a diffusion pattern different from the diffusion pattern obtained from the learning result is preselected.

As a result, in the diffusion pattern storage selection unit 415, M types of diffusion patterns (M=2) stored in each channel can be pre-selected according to the size of the adaptive gain, so that the number of diffusion pattern combinations can be greatly reduced. As a result, there is no need to calculate distortion for all combination numbers of the diffusion pattern, and vector quantization processing of noise source information can be efficiently performed with a small number of calculations.

Furthermore, the shape of the random code vector is pulse-shaped when the adaptive gain value is large (when the voice is strong), and is random when the adaptive gain value is small (when the voice is weak). Therefore, random code vectors with appropriate shapes can be used for the voiced area and the unvoiced area of the voice signal, so that the quality of the synthesized voice can be improved.

For simplicity of description, this embodiment is limited to the case where the number N of channels of the pulse vector generation unit is 3, and the number M of types of diffusion pulses stored in each channel in the diffusion mode storage and selection unit is 2. However, when the number of channels of the pulse vector generation unit and the number of diffusion patterns per channel in the diffusion pattern storage selection unit are different from those described above, the same effects and actions can be obtained.

For simplicity of description, in this embodiment, among the M kinds of diffusion patterns stored in each channel (M=2), one is the diffusion pattern obtained by the above-mentioned learning, and the other is a random pattern. However, if each channel stores at least one diffusion pattern obtained by learning in advance, even if it is not the case as above, the same effect and action can be expected.

In the present embodiment, the case where the adaptive code gain size information is used as a means for preselecting the diffusion pattern is described, but further effects can be expected to be obtained by using parameters other than the adaptive gain size information that represent the short-term characteristics of the voice signal. and function.

Furthermore, by constructing a speech signal communication system or a speech signal recording system having the above-mentioned speech signal encoder, the functions and effects of the sound source vector generation device described in Embodiment 1 can be obtained.

Moreover, this embodiment has explained the method of preselecting the diffusion mode by using the ideal adaptive sound source gain of the current processing frame that can be referred to when the noise source information is quantized, but instead of using the ideal adaptive sound source gain of the current frame The same structure can be adopted when decoding the adaptive sound source gain calculated immediately before the previous frame, and the same effect can also be obtained at this time.

(fourth embodiment)

Fig. 7 is a functional block diagram of a CELP type speech signal encoder according to this embodiment. In the present embodiment, in the CELP type speech signal encoder employing the sound source vector generation device of the first embodiment in the noise codebook, information available at the time of vector quantization of the noise source information is used to preselect the multiplicity of the diffusion patterns stored in the selection unit. a diffusion pattern. As a criterion for this preselection, it is characterized in that the magnitude of coding distortion (indicated by the S/N ratio) generated when the adaptive codebook number is specified is used.

Except for the peripheral part of the random codebook, it is the same as that of the CELP type speech signal encoder in Fig. 4 . Therefore, this embodiment describes in detail the vector quantization processing of noise source information.

As shown in FIG. 7, the CELP type speech signal encoder of this embodiment has an adaptive codebook 507, an adaptive code gain weighting unit 509, a noise codebook 508 composed of the sound source vector generation device described in the first embodiment, Noise code gain weighting unit 510, synthesis filter 505, distortion calculation unit 506, code number regulation unit 513, diffusion mode storage selection unit 515, pulse vector generation unit 516, pulse vector diffusion unit 517, diffusion vector adder 518 and distortion power Judgment unit 519.

However, in this embodiment, at least one of the M kinds of diffusion patterns (M≥2) stored in the diffusion pattern storage and selection unit 515 is a random pattern.

For ease of description, in the present embodiment, the number of channels N of the pulse vector generation unit is 3, and the number M of types of diffusion patterns of each channel stored in the diffusion pattern storage selection unit is 2, and it is assumed that in M kinds of diffusion patterns (M=2 ) one is a random mode, and the other is pre-learned, the diffusion mode obtained from the learning result after the quantization distortion generated by the noise source information vector quantization is small.

In the CELP type speech signal encoder of FIG. 7, the process of specifying the adaptive codebook number is performed before the noise source information vector quantization process. Therefore, at the time of vector quantization processing of the noise source number, the vector number of the adaptive codebook (adaptive code number), the ideal adaptive code gain (determined temporarily), and the target vector for adaptive codebook search can be referred to. In this embodiment, the preselection of the diffusion pattern is performed using adaptive codebook coding distortion (expressed by S/N ratio) which can be calculated from the above three kinds of information.

Specifically, immediately after the adaptive codebook search is completed, the adaptive code number and the value of the adaptive code gain (ideal gain) held by code number specifying section 513 are output to distortion calculating section 506 . Distortion calculating section 506 calculates the encoding distortion (S /N ratio). The calculated S/N ratio is output to distortion power determination section 519 .

The distortion power determination unit 519 first compares the S/N ratio received from the distortion calculation unit 506 with a preset threshold value. Next, the distortion power determination unit 519 sends the pre-selected control signal to the diffuse pattern storage selection unit 515 according to the result of the above magnitude comparison. When the content of the control signal is greater than the S/N ratio in the above-mentioned size comparison, it indicates that pre-learning is selected so that the encoding distortion generated by encoding the target vector for random codebook retrieval is small, and the resultant diffusion pattern is described above. A small S/N ratio in size comparison indicates that a random pattern of diffusion patterns is selected.

As a result, in the diffusion pattern storage selection unit 515, only one of the M diffusion patterns (M=2) stored in each channel is preselected, and the combination of diffusion patterns can be greatly reduced. Therefore, it is not necessary to calculate distortions for all combination numbers of the diffusion pattern, and it is possible to efficiently specify random code numbers with a small number of calculations. Furthermore, the shape of the random code vector is pulse-like when the S/N ratio is large, and random-like when the S/N ratio is small. Therefore, the shape of the random code vector can be changed according to the short-term characteristics of the speech signal, thereby improving the quality of the synthesized speech.

For simplicity of description, this embodiment is limited to the case where the number N of channels of the pulse vector generation unit is 3, and the number M of types of diffusion pulses stored in each channel in the diffusion mode storage and selection unit is 2. However, even when the number of channels of the pulse vector generation unit and the type of diffusion pattern per channel are different from those described above, the same effect and action can be obtained.

For simplicity of description, this embodiment will also describe the case where among the M types of diffusion patterns stored in each channel (M=2), one is the diffusion pattern acquired by the above-mentioned learning, and the other is a random pattern. However, if the diffusion pattern of at least one random pattern is stored in advance for each channel, the same effects and operations can be expected even in cases other than the above.

In this embodiment, although only the size information of the coding distortion (expressed by S/N ratio) generated by the specified adaptive code number is used as the means for pre-selecting the diffusion mode, if the information that further accurately expresses the short-term characteristics of the voice signal is also used , can be expected to have further effects and functions.

Furthermore, by constructing a speech signal communication system or a speech signal recording system having the above-mentioned speech signal encoder, the functions and effects of the sound source vector generation device described in the first embodiment can be obtained.

Fifth Embodiment

Fig. 8 is a functional block diagram of a CELP type speech signal encoder according to a fifth embodiment of the present invention. In this CELP type speech signal encoder, LPC coefficients are obtained by performing autocorrelation analysis and LPC analysis on the input speech data 601 in the LPC analysis unit 600 . Further, while encoding the obtained LPC coefficients to obtain LPC codes, the obtained LPC codes are decoded to obtain decoded LPC coefficients.

Then, in the sound source generating unit 602, take out the sound source sampling stored in the adaptive codebook 603 and the random codebook 604 (referred to as adaptive code vector (or adaptive sound source) and random code vector (or noise source)), and respectively sent to the LPC synthesis unit 605.

In the LPC synthesis unit 605, the two sound sources obtained by the sound source generation unit 602 are filtered using the decoded LPC coefficients obtained by the LPC analysis unit 600, thereby obtaining two synthesized voices.

In the comparison unit 606, analyze the relationship between the two synthesized voices obtained by the LPC synthesis unit 605 and the input voice 601, find the optimum value (optimum gain) of the two synthesized voices, and use the optimum gain to adjust the power of each The synthesized voices are added to obtain the total synthesized voice, and the distance between the total synthesized voice and the input voice is calculated.

Again, for the sampling of all sound sources of the adaptive codebook 603 and the random codebook 604, calculate the distance between a plurality of synthesized voices obtained by driving the sound source generation unit 602 and the LPC synthesis unit 605 and the input voice 601, and calculate the distance between the obtained results The minimum hour audio sample index number.

Send the obtained optimal gain, the sound source sampling index number and the two sound sources corresponding to the index number to the parameter encoding unit 607 . After the parameter coding unit 607 obtains the gain code by performing optimal gain coding, the LPC code and the sound source sample index number are combined and sent to the transmission line 608 .

Based on the two sound sources corresponding to the gain code and the index number, an actual sound source signal is generated, the signal is stored in the adaptive codebook 603, and old sound source samples are discarded.

Furthermore, in the LPC synthesis unit 605, an auditory weighting filter is usually used as well, and the filter uses a linear prediction coefficient, a high-frequency enhancement filter and a long-term prediction coefficient (obtained by long-term prediction analysis of the input voice). In general, an audio source search for an adaptive codebook and a random codebook is performed using a subdivided section (referred to as a subframe) of the analysis section.

Next, this embodiment will describe in detail the LPC coefficient vector quantization in the LPC analysis unit 600 .

FIG. 9 shows a functional block diagram for realizing the vector quantization algorithm executed in the LPC analysis unit 600 . The vector quantization block shown in FIG. 9 includes a target extraction unit 702 , a quantization unit 703 , a distortion calculation unit 704 , a comparison unit 705 , a decoded vector storage unit 707 and a vector smoothing unit 708 .

In target extracting section 702 , a quantization target is calculated from input vector 701 . The method of extracting the target will now be described in detail.

The input vector in this embodiment is composed of two types of vectors: a parameter vector obtained by analyzing the encoding target frame; and a parameter vector obtained by similarly analyzing a future frame. The target extraction unit 702 calculates the quantization target by using the input vector and the decoded vector of the previous frame stored in the decoded vector storage unit 707 . Equation (8) shows an example of a calculation method.

X(i)＝{S _t (i)+p(d(i)+S _t+1 (i)/2}/(1+p)(8)

X(i): target vector

i: vector element number

S _t (i), S _t+1 (i): input vector

t: time (frame number)

p: weighting coefficient (fixed)

d(i): previous frame decoding vector

The idea of the above object extraction method is shown below. In typical vector quantization, the parameter vector S _t (i) of the current frame is used as the target X(i), and is fitted by formula (9).

$\mathrm{<span\; class="notranslate">\; En</span>}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; I</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; Cn</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 9</span>}<span\; class="notranslate">\; )</span>$

En: the distance from the n number vector

X(i): quantitative target

Cn(i): code vector

n: code vector number

i: vector dimension

I: vector length

Therefore, in the conventional vector quantization, coding distortion still leads to degradation of sound quality. This becomes a big problem in ultra-low bit-rate coding in which a certain degree of coding distortion cannot be avoided even if measures such as vector predictive quantization are taken.

Therefore, in the present embodiment, the midpoint of the previous and subsequent decoded vectors is focused on as a direction in which errors are difficult to be detected audibly, and the decoded vector is derived there, thereby improving the audible sense. This is due to the fact that when the parameter vector interpolation characteristic is good, it is difficult to hear the deterioration of the time-continuous characteristic aurally. This is explained below with reference to FIG. 10 showing a vector space.

First, the decoding vector of the previous frame is d(i), and the future parameter vector is S _t+1 (i) (in fact, it is better to be the future decoding vector, but it cannot be coded in the current frame, so the parameter vector is used instead), then Code vector Cn(i): (1) is code vector Cn(i): (2) is closer to parameter vector S _t (i), but in fact Cn(i): (2) is close to d(i) and S On the connection of _t+1 (i), it is less likely to hear degradation than Cn(i): (1). Therefore, using this characteristic, if the target X(i) is taken as a vector from S _t (i) to a certain extent close to the midpoint of d(i) and S _t+1 (i), then the decoding The vector leads to a direction with less audible distortion.

In this embodiment, such target movement can be realized by introducing the estimation of the following formula (10).

X(i)＝{S _t (i)+p(d(i)+S _t+1 (i)/2}/(1+p)(10)

X(i): quantization target vector

i: vector element number

S _t (i), S _t+1 (i): input vector

t: time (frame number)

p: weighting coefficient (fixed)

d(i): previous frame decoding vector

The first half of formula (10) is the general vector quantization estimation formula, and the second half is the auditory weighting component. In order to use the above estimation formula for quantification, the estimation formula is differentiated at each X(i), and the result of the differentiation is set to 0, then formula (8) can be obtained.

The weighting coefficient P is a positive constant. When its value is 0, it is the same as general vector quantization. When it is infinite, the target is completely located at the midpoint. If P is very large, the target will greatly deviate from the parameter vector S _t (i) of the current frame, and the auditory clarity will decrease. According to the audition experiment of the decoded voice signal, it is confirmed that good performance is obtained when 0.5<p<1.0.

The quantization unit 703 quantizes the quantization target obtained by the target extraction unit 702 to obtain a vector code and simultaneously obtain a decoding vector, and send it to the distortion calculation unit 704 together with the vector code.

In this embodiment, predictive vector quantization is adopted as a quantization method. Next, predictive vector quantization will be described.

FIG. 11 shows a functional block diagram of predictive vector quantization. Predictive vector quantization is an algorithm for predicting using vectors (synthetic vectors) obtained by encoding and decoding in the past, and performing vector quantization on the prediction errors.

A vector codebook 800 storing center samples (code vectors) of a plurality of prediction error vectors is generated in advance. Usually, the codebook is generated by using the LBG algorithm (IEEE TRANSACTIONSON COMMUNICATIONS, VOL.COM-28, NO.1, pp84-95, JANUARY 1980) based on multiple vectors obtained by analyzing multiple voice data.

The vector 801 of the quantization target is predicted in the prediction unit 802 . The prediction is performed using the past composite vectors stored in the state storage unit 803 , and the obtained prediction error vector is sent to the distance calculation unit 804 . Here, as an aspect of prediction, prediction using a fixed coefficient when the number of times of prediction is one is mentioned. The following equation (11) shows the prediction error vector calculation equation at the time of the above prediction.

Y(i)=X(i)-βD(i)(11)

Y(i): prediction error vector

X(i): quantitative target

β: prediction coefficient (scalar)

D(i): synthetic vector of the previous frame

i: vector dimension

The value of the prediction coefficient β in the above formula is generally 0<β<1.

In the distance calculation unit 804 , the distance between the prediction error vector obtained by the prediction unit 802 and the code vector stored in the vector codebook 800 is calculated. Equation (12) below shows this distance calculation formula.

$\mathrm{<span\; class="notranslate">\; En</span>}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; I</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; Cn</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 12</span>}<span\; class="notranslate">\; )</span>$

En: the distance from the n number vector

T(i): prediction error vector

Cn(i): code vector

n: code vector number

i: vector dimension

I: vector length

The distances to the respective code vectors are compared in the retrieval section 805 , and the number of the code vector with the smallest distance is output as a vector code 806 . That is, control the vector codebook 800 and the distance calculation unit 804 to find the number of the codevector with the smallest distance among all the codevectors stored in the vector codebook 800 , and use the vector number as the vector code 806 .

Furthermore, according to the final vector code, vector decoding is performed using the code vector obtained from the vector codebook 800 and the past decoded vector stored in the state storage unit 803, and the content of the state storage unit 803 is updated with the obtained composite vector. Therefore, the vector decoded here can be used for prediction in the next decoding.

Decoding of the above-mentioned example of the prediction mode (the number of predictions is 1, and the coefficient is fixed) is performed using the following equation (13).

Z(i)=CN(i)+βD(i)(13)

Z(i): decoding vector (used as D(i) in the next encoding)

N: vector encoding

CN(i): code vector

β: prediction coefficient (scalar)

D(i): synthetic vector of the previous frame

I: vector dimension

On the other hand, in the decoder, decoding is performed by finding a code vector from the transmitted vector code. The decoder is pre-prepared with the same vector codebook and state storage unit as the encoder, and uses the algorithm with the same function as the retrieval unit decoder in the above encoding algorithm to perform decoding. The above is the vector quantization performed at the quantization unit 703 .

In the distortion calculation unit 704 , the auditory weighted coding distortion is calculated according to the decoded vector obtained by the quantization unit 703 , the input vector 701 and the decoded vector of the previous frame stored in the decoded vector storage unit 707 . The following formula (14) shows a calculation formula.

Ew＝∑(V(i)-S _t (i)) ² +p{V(i)-(d(i)+S _t+1 (i)/2} ² (14)

Ew: weighted encoding distortion

S _t (i), S _t+1 (i): input vector

t: time (frame number)

i: vector element number

V(i): decoded vector

p: weighting coefficient (fixed)

d(i): previous frame decoding vector

In Equation (14), the weighting coefficient p is the same as the coefficient of the object calculation equation used by the object extracting unit 702 . Send the above weighted coding distortion value, decoded vector and vector code to the comparison unit 705 .

Comparing section 705 sends the vector code sent from distortion calculating section 704 to transmission line 608, and uses the decoded vector sent from distortion calculating section 704 to update the contents of decoded vector storage section 707.

According to the above-mentioned embodiment, the target vector is corrected in the target extracting unit 702 to a vector at a position close to the midpoint of d(i) and S _t+1 (i) from S _t (i) to some extent, so weighting can be performed. Retrieve without feeling aurally degraded.

So far, it has been described that the present invention is applicable to the low bit rate speech signal encoding technique used in portable telephones and the like, but the present invention can be applied not only to speech signal encoding but also to music encoders. Parametric vector quantization with better interpolation in image coders.

The LPC encoding of the LPC analysis unit in the above algorithm is usually transformed into a general LSP (line spectrum pair) and other parameter vectors that are convenient for encoding, and vector quantization (VQ) is performed using Euclidean distance and weighted Euclidean distance.

In this embodiment, the object extraction unit 702 is controlled by the comparison unit 705, and sends the input vector 701 to the vector smoothing unit 708, and the object extraction unit 703 receives the modified input vector in the vector smoothing unit 708, and then extracts the object.

At this time, the weighted coding distortion value sent from the distortion calculation section 704 is compared in the comparison section 705 with the reference value prepared inside the comparison section. According to the result of this comparison, the processing is divided into two types.

When the reference value is not reached, the vector code sent from the distortion calculation unit 704 is sent to the transmission line 606, and the content of the decoded vector storage unit 707 is updated with the decoded vector sent from the distortion calculation unit 704. This update is performed by overwriting the content of the decoded vector storage unit 707 with the obtained decoded vector. Then, transition to the next frame parameter encoding process.

On the contrary, when it is above the reference value, control the vector smoothing unit 708 to modify the input vector, so that the target extraction unit 702, the quantization unit 703 and the vector calculation unit 704 function again to perform re-encoding.

Until the comparison section 705 reaches the reference value, the encoding process is repeated. However, sometimes the reference value cannot be reached after repeated several times. Therefore, the comparison unit 705 has a counter inside to calculate the number of times it is judged to be above the reference value. Time processing and counter clearing.

In the vector smoothing unit 708, receiving the control of the comparison unit 705, according to the input vector obtained by the target extraction unit 702 and the previous frame decoding vector obtained from the decoding vector storage unit 707, the following formula (15) is used to modify as one of the input vectors The current frame parameter vector S _t (i), and the modified input vector is sent to the target extraction unit 702 .

S _t (i)←(1-q) S _t (i)+q(d(i)+S _t+1 (i))/2(15)

The above q is a smoothing coefficient, indicating how close the current frame parameter vector is to the midpoint between the previous frame decoding vector and the future frame parameter vector. According to the coding implementation, it is confirmed that good performance can be obtained when 0.2<q<0.4 and the upper limit of the number of internal iterations of the comparison unit 705 is 5-8 times.

In this embodiment, although the quantization section 703 employs predictive vector quantization, it is highly likely that the weighted encoding distortion obtained by the distortion calculation section 704 will be reduced by the smoothing process described above. The reason for this is to make the quantization target closer to the previous frame decoded vector by smoothing. Therefore, by repeating the encoding controlled by comparing section 705 , it is more likely that the distortion comparison by comparing section 705 will not reach the reference value.

In the decoder, a decoding unit corresponding to the quantization unit of the encoder is prepared in advance, and decodes based on the vector code sent from the transmission line.

This embodiment is also used for LSP parameter quantization (quantization unit predicts VQ) in CELP type coding to perform coding and decoding experiments of voice signals. As a result, it was confirmed that the auditory sound quality can be improved as a matter of course, and the objective value (S/N ratio) can also be improved. This is because the iterative encoding process with vector smoothing achieves the effect of suppressing the predicted VQ encoding distortion even when the frequency spectrum changes rapidly. Conventional predictive VQ has a disadvantage in that the spectral distortion in the part where the frequency spectrum changes rapidly, such as the part where the speech starts, becomes rather large due to the prediction based on the past composite vector. However, applying this embodiment, when the distortion is large, the smoothing process is performed until the distortion becomes smaller. Therefore, although the target deviates from the actual parameter vector, the encoding distortion becomes smaller, and the overall degradation of the speech signal is reduced when decoding. Therefore, according to the present embodiment, not only the sound quality can be improved perceptually but also the objective value can be improved.

In this embodiment, the characteristics of the comparison unit and the vector smoothing unit can be used to control the direction of its deterioration in a direction that is relatively imperceptible to the auditory sense when the vector quantization distortion is large, and when the quantization unit adopts predictive vector quantization by Repeatedly performing smoothing + encoding until the encoding distortion becomes small can also increase the objective value.

So far, it has been explained that the present invention is applicable to low-bit-rate speech coding techniques used in portable telephones, etc., but the present invention is not only for speech signal coding, but also can be used for parameter vectors with better interpolation in music coders and image coders. Quantify.

(sixth embodiment)

Next, a CELP type speech signal encoder according to a sixth embodiment of the present invention will be described. This embodiment is the same as the above-mentioned fifth embodiment except that the quantization method adopts the quantization algorithm of the quantization unit of multi-stage predictive vector quantization. That is, the noise source vector generator of the first embodiment described above is used as the random codebook. The quantization algorithm of the quantization unit will now be described in detail.

A functional block diagram of the quantization unit is shown in FIG. 12 . In multi-level vector quantization, after the target vector quantization is performed, the quantized target codeword is decoded using its codebook, and the difference between the encoded vector and the original target (called the encoding distortion vector) is calculated, and then the obtained encoding distortion Vectors are quantized.

The vector codebook 899 and the vector codebook 900 storing a plurality of prediction error vector center samples (code vectors) are generated in advance. These codebooks are generated by applying the same algorithm as a typical "multi-level vector quantization" codebook generation method to a plurality of prediction error vectors for learning. That is, usually based on a plurality of vectors obtained by analyzing many voice data, the LBG algorithm (IEEE TRANSACTIONS ON COMMUNICATIONS, VOL.COM-28, NO.1, pp84-95, JANUARY 1980) is used to generate the above codebook. However, the total learning of the vector codebook 899 is a collection of many quantization targets, and the total learning of the vector codebook 900 is a collection of encoding distortion vectors when the quantization codebook 899 is used to encode the above-mentioned many quantization targets.

First, the quantization target vector 901 is predicted in the prediction unit 902 . The prediction is performed using the past composite vector stored in the state storage unit 903 , and the obtained prediction error vector is sent to the distance calculation unit 904 and the distance calculation unit 905 .

In the present embodiment, as a prediction form, prediction using a fixed coefficient is given when the number of predictions is one. The following equation (16) shows the prediction error vector calculation equation when such prediction is used.

Y(i)=X(i)-βD(i)(16)

Y(i): prediction error vector

X(i): quantitative target

β: prediction coefficient (scalar)

D(i): synthetic vector of the previous frame

i: vector dimension

In the above formula, the value of the prediction coefficient β is usually 0<β<1.

In the distance calculation unit 904 , the distance between the prediction error vector obtained by the prediction unit 902 and the code vector A stored in the vector codebook 899 is calculated. The following equation (17) shows the distance calculation equation.

$\mathrm{<span\; class="notranslate">\; En</span>}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; I</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; <figure-callout\; id="1"\; label="C"\; filenames="C9880155600371.png,C9880155600461.png"\; state="\{\{state\}\}">C</figure-callout></span>}\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 17</span>}<span\; class="notranslate">\; )</span>$

En: the distance from n number vector A

X(i): prediction error vector

C1n(i): code vector A

n: number of code vector A

i: vector dimension

I: vector length

In the search section 906, the distances to the respective code vectors A are compared, and the number of the code vector A with the smallest distance is used as the code of the code vector A. That is, control the vector codebook 899 and the distance calculation unit 904 to obtain the number of the codevector A with the smallest distance among all the codevectors stored in the vector codebook 899, and use this number as the code of the codevector A. Then, the code of the code vector A and the decoded vector A obtained from the vector codebook 899 referring to the code are sent to the distance calculating section 905 . The encoding of the code vector A is sent to the transmission line and the retrieval unit 907.

The distance calculation unit 905 obtains the coding distortion vector from the prediction error vector and the decoded vector A obtained from the retrieval unit 906, or refers to the code of the code vector A obtained from the retrieval unit 906, obtains the magnitude from the magnitude storage unit 908, and then calculates the above code The distance between the distortion vector and the code vector B stored in the vector codebook 900 multiplied by the above amplitude is obtained, and the distance is sent to the retrieval unit 907 . The following equation (18) shows the distance calculation equation.

Z(i)=Y(i)-C1N(i)

$\mathrm{<span\; class="notranslate">\; Em</span>}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; I</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; Z</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; NC</span>}\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 18</span>}<span\; class="notranslate">\; )</span>$

Z(i): decoded distortion vector

Y(i): prediction error vector

C1N(i): decoded vector A

N: encoding of code vector A

Em: the distance from the m number vector B

aN and the magnitude corresponding to the encoding of the code vector A

C2m(i): code vector B

m: number of code vector B

i: vector dimension

I: vector length

In the search section 907, the distances to the respective code vectors B are compared, and the number of the code vector B with the smallest distance is used as the code of the code vector B. That is, control the vector codebook 900 and the distance calculation unit 905 to find the number of the codevector B with the smallest distance among all the codevectors B stored in the vector codebook 900, and use this number as the code of the codevector B. Then, codes of code vector A and code vector B are combined to form vector 909 .

The retrieval unit 907 also uses the decoding vectors A and B obtained from the vector codebook 899 and the vector codebook 900, the magnitude obtained from the magnitude storage unit 908, and the past decoding codes stored in the state storage unit 903 according to the coding of the code vectors A and B. The vector is decoded, and the content of the state storage unit 903 is updated with the obtained composite vector. (Therefore, the vector decoded here is used for prediction when the next encoding is performed.) Decoding in the prediction of this embodiment (the number of predictions is 1, and the coefficient is fixed) is performed using the following equation (19).

Z(i)＝C1N(i)+aN·C2M(i)+βD(i) (19)

Z(i): decoding vector (used as D(i) in the next encoding)

N: encoding of code vector A

M: encoding of code vector B

C1M: decode vector A

C2M: Decoding Vector B

aN: magnitude corresponding to encoding of code vector A

β: prediction coefficient (scalar)

D(i): synthetic vector of the previous frame

i: vector dimension

The width stored in the width storage unit 908 is set in advance, and the setting method is shown below. A lot of speech data is coded, and the total coding distortion of the following equation (20) is obtained for each code of the first-order code vector, and learning is performed to minimize the distortion to set the amplitude.

$\mathrm{<span\; class="notranslate">\; EN</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; I</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; Y</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; C</span>}\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; NC</span>}\mathrm{<span\; class="notranslate">\; 2</span>}{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 20</span>}<span\; class="notranslate">\; )</span>$

EN: Coding distortion when code vector A is coded as N

N: encoding of code vector A

t: the time when the encoding of the code vector A is N

Y _t (i): prediction error vector at time t

C1N(i): decoded vector A

aN: magnitude corresponding to encoding of code vector A

C2m _t (i): code vector B

m _t : number of code vector B

i: vector dimension

I: vector length

That is, after encoding, the distortion of the above-mentioned expression (20) is set and corrected so that the differential value in each amplitude is 0, thereby performing amplitude learning. Then, the above-described encoding+learning is repeated to find the optimum width.

On the other hand, in the decoder, decoding is carried out by calculating the code vector according to the transmitted vector code. The decoder has the same vector codebook (corresponding to code vectors A, B), amplitude storage unit and state storage unit as the encoder, and uses the same algorithm as the decoding function of the retrieval unit (corresponding to code vector B) in the above-mentioned encoding algorithm to decode.

Therefore, in this embodiment, the second-order code vector is adapted to the first-order code vector with a small amount of calculation by using the features of the amplitude storage unit and the distance calculation unit, so that the amplitude distortion can be reduced.

So far, it has been explained that the present invention is applicable to low-bit-rate speech signal encoding techniques used in portable telephones, etc., but the present invention is not only applicable to speech signal encoding, but can also be used in music encoders, image encoders, etc., which have better interpolation performance. Parameter vector quantization.

(seventh embodiment)

Next, a CELP type speech signal encoder according to a seventh embodiment of the present invention will be described. The aspect of the present invention is an example of an encoder that can reduce the amount of computation for time code retrieval using an ACELP type noise codebook.

Fig. 13 shows a functional block diagram of a CELP type speech coder according to this embodiment. In this CELP type voice signal encoder, the filter coefficient analysis unit 1002 performs linear predictive analysis on the input voice signal 1001 to obtain synthesis filter coefficients, and outputs the obtained synthesis filter coefficients to the filter coefficient quantization unit 1003 . Filter coefficient quantization section 1003 quantizes the input synthesis filter coefficients and outputs them to synthesis filter 1004 .

Synthesis filter 1004 is established based on filter coefficients supplied from filter coefficient quantization unit 1003 and is driven by excitation signal 1011 . The excitation signal 1011 is obtained by adding the result obtained by multiplying the adaptive vector 1006 output by the adaptive codebook 1005 by the adaptive gain 1007 and the result obtained by multiplying the noise vector 1009 output by the random codebook 1008 by the noise gain 1010 .

Here, the adaptive codebook 1005 is a codebook that stores a plurality of adaptive vectors obtained in the past for the excitation signal for the synthesis filter for each pitch cycle, and the noise codebook 1007 is a codebook that stores a plurality of noise vectors. As the random codebook 1007, the excitation vector generator of the first embodiment described above can be used.

Distortion calculating section 1013 calculates distortion between synthesized speech signal 1012 which is an output of synthesis filter 1004 driven by excitation signal 1011 and input speech signal 1001, and performs code search processing. The code search process is a method of specifying the number of the adaptive vector 1006 and the number of the noise vector 1009 used to minimize the distortion calculated by the distortion calculation unit 1013, and calculating the optimum values of the adaptive gain 1007 and the noise gain 1010 multiplied by each output vector processing.

The coded output unit 1014 outputs the adaptive vector 1006 and the adaptive vector 1009 selected in the distortion calculation unit 1013 by multiplying the filter coefficient quantization value obtained from the filter coefficient quantization unit 1003 and the noise vector 1009 respectively. Gain 1007 and Noise Gain 1009 are encoded results. The information output from the code output unit 1014 is transmitted or stored.

In the code retrieval process in the distortion calculation unit 1013, usually, the adaptive codebook component in the excitation signal is retrieved first, and then the noise codebook component in the excitation signal is retrieved.

The retrieval of the above-mentioned noise components uses the orthogonal retrieval described below.

In the orthogonal search, the noise vector c that maximizes the search reference value Eort (=Nort/Dort) of the equation (21) is specified.

$\mathrm{<span\; class="notranslate">\; Eort</span>}<span\; class="notranslate">\; (</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; Nort</span>}}{\mathrm{<span\; class="notranslate">\; Dort</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{{<span\; class="notranslate">\; [</span><span\; class="notranslate">\; \{</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; P</span>}}^{\mathrm{<span\; class="notranslate">\; t</span>}}{\mathrm{<span\; class="notranslate">\; h</span>}}^{\mathrm{<span\; class="notranslate">\; t</span>}}\mathrm{<span\; class="notranslate">\; Hc</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}^{\mathrm{<span\; class="notranslate">\; t</span>}}\mathrm{<span\; class="notranslate">\; HP</span>}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; HP</span>}<span\; class="notranslate">\; \}</span>\mathrm{<span\; class="notranslate">\; Hc</span>}<span\; class="notranslate">\; ]</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; c</span>}}^{\mathrm{<span\; class="notranslate">\; t</span>}}{\mathrm{<span\; class="notranslate">\; h</span>}}^{\mathrm{<span\; class="notranslate">\; t</span>}}\mathrm{<span\; class="notranslate">\; Hc</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; p</span>}}^{\mathrm{<span\; class="notranslate">\; t</span>}}{\mathrm{<span\; class="notranslate">\; h</span>}}^{\mathrm{<span\; class="notranslate">\; t</span>}}\mathrm{<span\; class="notranslate">\; HP</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; p</span>}}^{\mathrm{<span\; class="notranslate">\; t</span>}}{\mathrm{<span\; class="notranslate">\; h</span>}}^{\mathrm{<span\; class="notranslate">\; t</span>}}\mathrm{<span\; class="notranslate">\; Hc</span>}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; twenty\; one</span>}<span\; class="notranslate">\; )</span>$

Nort: Molecular term of Eort

Dort: the denominator term of Eort

p: specified adaptation vector

H: matrix of synthetic filter coefficients

H ^t : transpose matrix of H

X: target signal (the result of the difference between the input speech signal and the zero-input response of the synthesis filter)

c: noise vector

The orthogonal search is a search method in which noise vectors which are predetermined as candidates for adaptive vectors are orthogonalized respectively, and one of the plurality of orthogonal noise vectors is specified with the least distortion. Compared with the non-orthogonal search, this search method is characterized in that it can improve the accuracy of the specified noise vector, thereby improving the quality of the synthesized speech signal.

In the ACELP method, only a few pulses with polarity are used to form the noise vector. Using this point, the numerator term (Nort) of the search reference value shown in the formula (21) is converted into the following formula (22), thereby reducing the calculation of the numerator term.

Nort＝{a ₀ ψ(l ₀ )+a ₁ ψ(l ₁ )+…a _n-1 ψ(l _n-1 )} ² (22)

a _i : Polarity of the i-th pulse (+1/-1)

l _i : the position of the i-th pulse

N: number of pulses

ψ: {(p ^t H ^t Hp)x-(x ^t Hp)Hp}H

The value of ψ in formula (22) is pre-calculated as a pre-processing and expanded in the array, then the (N-1) elements in the array ψ can be added with signs, and the result is squared to calculate Molecular term of formula (21).

The distortion calculation unit 1013 that can reduce the amount of computation for the denominator term will be described in detail below.

A functional block diagram of the distortion calculation unit 1013 is shown in FIG. 14 . The speech signal encoder in this embodiment has a configuration in which adaptive vector 1006 and noise vector 1009 are input to distortion calculation section 1013 in the configuration shown in FIG. 13 .

In FIG. 14 , the following three types of processing are performed as pre-processing for calculating the distortion of the input noise vector.

(1) Calculate the first matrix (N): Calculate the power (p ^t H ^t Hp) of the vector obtained after the synthesis filter synthesizes the adaptive vector and the autocorrelation matrix (H ^t H) of the filter coefficients in the synthesis filter, A matrix N (=( ^pt H ^t Hp)H ^t H) is calculated by multiplying the above-mentioned power by each element of the above-mentioned autocorrelation matrix.

(2) Calculate the second matrix (M): synthesize the vectors obtained after the synthesis filter synthesizes the adaptive vector in counterclockwise order, and take the vector product of the resulting signal (p ^t H ^t H) to calculate Matrix M.

(3) Generation of the third matrix (L): The matrix N calculated in (1) and the matrix M calculated in (2) are differentiated to generate a matrix L.

Also, the denominator term (Dort) of Equation (21) can be expanded into Equation (23).

Dort＝(c ^t H ^t Hc)(p ^t H ^t Hp)-(p ^t H ^t Hc) ² (23)

＝c ^t Nc-(r ^t c) ²

＝c ^t Nc-(r ^t c) ^t (r _t c)

＝c ^t Nc-(c ^t rr ^t c)

＝c ^t Nc-(c ^t Mc)

＝c ^t (NM)c

=c ^t Lc

N: (p ^t H ^t Hp) H ^t H ← above preprocessing (1)

r: p ^t H ^t H ← the above preprocessing (2)

M: rr ^t ← the above preprocessing (2)

L: N-M ← above pre-processing (3)

c: noise vector

Thus, by replacing the calculation method of the denominator term (Dort) when calculating the search reference value (Eort) in Equation (21) with Equation (23), the random codebook component can be specified with a small amount of computation.

The denominator term is calculated using the matrix L and the noise vector 1009 obtained from the above pre-processing.

Here, for the sake of simplicity, the sampling frequency of the input voice signal is 8000Hz, the unit time width (frame time) of the algebraic structure noise codebook retrieval is 10ms, and the noise vector is combined with the rule of 5 unit pulses (+1/-1) every 10ms In the case of generation, the calculation method of the denominator term based on the formula (23) will be described.

It is also assumed that the five unit pulses constituting the noise vector are composed of pulses at one of the positions specified in Group 0 to Group 4 shown in Table 2, and the candidate noise vector can be described by the following formula (24).

C＝a ₀ δ(kl ₀ )+a ₁ δ(kl ₁ )+…+a ₄ δ(kl ₄ ) (24)

(k=0, 1, ... 79)

a _i : the polarity of the pulse belonging to the i group (+1/-1)

l _i : the position of the pulse belonging to the i group

Table 2

Group No symbol alternate pulse position

Number 0 ±1 0, 10, 20, 30, ..., 60, 70 1 ±1 2, 12, 22, 32, ..., 62, 72 2 ±1 2, 16, 26, 36, ..., 66, 76 3 ±1 4, 14, 24, 34, ..., 64, 74 4 ±1 8, 18, 28, 38, ..., 68, 78

In this case, the denominator term (Dort) shown in the formula (23) can be obtained by the following formula (25).

$\mathrm{<span\; class="notranslate">\; Dort</span>}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; 4</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; 4</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}\mathrm{<span\; class="notranslate">\; L</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; l</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; l</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 25</span>}<span\; class="notranslate">\; )</span>$

a _i : the polarity of the pulse belonging to the i group

l _i : the position of the pulse belonging to the i group

L(l _i , l _j ): the elements of l _i row and l _j column in matrix L

According to the above description, it is proved that when ACELP type noise codebook is used, formula (22) can be used to calculate the numerator term (Nort) of the code retrieval reference value of formula (21), and formula (25) can be used to calculate its denominator term (Dort). Therefore, when using ACELP type noise codebook, instead of calculating the reference value of formula (21) as it is, the numerator and denominator terms are calculated by using (22) and formula (25), which can greatly reduce the amount of code retrieval operations .

In the present embodiment described above, random codebook retrieval without preselection has been described. However, the present invention is applied in the case of preselecting a noise vector having a large value of Equation (22), and calculating Equation (21) for noise vectors converged to a plurality of candidates by preselection, and selecting a noise vector having the largest value. can also achieve the same effect.

Claims (28)

1. A diffusion vector generator for use in an excitation vector generator for a CELP-type speech signal encoder/speech signal decoder, comprising:

a pulse vector generating unit that generates a pulse vector having a pulse with a polarity unit on a certain element of a vector axis;

a diffusion pattern storage unit storing a plurality of diffusion patterns;

a selection unit that selects one diffusion pattern from the plurality of diffusion patterns stored in the diffusion pattern storage unit; and

and a pulse vector diffusion unit which performs convolution operation of the selected diffusion pattern and the pulse vector to generate a diffusion vector.

2. The diffusion vector generation apparatus according to claim 1, wherein said excitation vector generation apparatus generates excitation vectors from N diffusion vectors according to the following equation

c: sound source vector

ci: diffusion vector

i: channel number (i = 1-N)

n: vector element numbers (n =0 to L-1, where L is the length of the excitation vector).

3. The diffusion vector generation apparatus of claim 1, wherein the pulse vector is generated by an algebraic structure noise codebook.

4. A CELP-type speech signal encoder, the CELP-type speech signal encoder comprising:

diffusion vector generating means for the excitation vector generating means;

a noise codebook for vector-quantizing the noise source information;

a synthesis filter for generating a synthesized speech using the excitation vector outputted from the excitation vector generation device as a noise code vector;

a distortion calculator that calculates quantization distortions of the generated synthesized speech and the input speech;

a switching unit that switches a combination of a pulse position and a pulse polarity and a diffusion pattern constituting a pulse vector;

a decision unit that decides a combination of a pulse position, a pulse polarity, and a diffusion pattern at which the quantization distortion calculated by the distortion calculator is minimum, and determines a noise codebook number; and

an adaptive codebook storing adaptive codevectors representing tonal components of an input voice;

the diffusion vector generation device includes:

a pulse vector generating unit that generates a pulse vector having a pulse with a polarity unit on a certain element of a vector axis;

a diffusion pattern storage unit storing a plurality of diffusion patterns;

a selection unit that selects one diffusion pattern from the plurality of diffusion patterns stored in the diffusion pattern storage unit; and

and a pulse vector diffusion unit which performs convolution operation of the selected diffusion pattern and the pulse vector to generate a diffusion vector.

5. The CELP-type speech signal encoder according to claim 4, wherein a diffusion pattern obtained by learning in advance and having less quantization distortion generated when quantizing the noise source information vector is stored in a storage unit in the excitation vector generation device.

6. The CELP-type speech signal encoder of claim 5, wherein the storage unit in the excitation vector generation apparatus stores at least one diffusion pattern obtained by learning for each channel.

7. The CELP speech signal encoder of claim 6, wherein the diffusion mode derived from learning is selected when a gain value of the code adaptive codebook is greater than a predetermined threshold value.

8. The CELP-type speech signal encoder of claim 4, wherein at least one of the diffusion patterns stored in the storage unit of the excitation vector generation apparatus in each channel is a random pattern.

9. The CELP-type speech signal encoder according to claim 4, wherein at least one of the diffusion patterns stored in the storage unit in the excitation vector generator in each channel is a diffusion pattern obtained by learning in advance and having a smaller quantization distortion generated when quantizing the noise source information vector, and at least one of the diffusion patterns is a random pattern.

10. The CELP-type speech signal encoder of claim 9, wherein the diffusion vector of the random pattern is selected in case the coding distortion generated in the decision of the adaptive codebook number is larger than a predetermined threshold.

11. The CELP-type speech signal encoder of claim 4, wherein M from a desired diffusion mode ^N In the full combination, a combination number representing a combination of diffusion patterns selected for each channel is determined to minimize quantization distortion generated when a noise source information vector is quantized.

12. The CELP-type speech signal encoder according to claim 11, wherein a combination of the diffusion patterns is preselected using the speech parameters obtained in advance, quantization distortion generated when the noise source information is vector-quantized is minimized, and a combination number indicating the combination of the diffusion patterns selected for each channel is determined from the combination of the diffusion patterns obtained by preselection.

13. The CELP-type speech signal encoder of claim 12, wherein the combination of preselected dispersion patterns is switched according to the speech interval analysis results.

14. The CELP-type speech signal encoder of claim 4, further comprising:

a target extraction unit for calculating a quantization target vector by using a parameter vector of a speech parameter obtained by analyzing a current encoding target frame, a parameter vector obtained by analyzing a frame in the future from the encoding target frame, and a decoding vector of a frame in the past from the current encoding target frame; and

and a vector quantization unit for encoding the calculated quantization target vector to obtain a noise codebook number of a current encoding target frame.

15. The CELP-type speech signal encoder of claim 14, wherein the target extraction unit computes a quantization target vector according to

X(i)＝{S _t (i)+p(d(i)+S _t+1 (i))/2}/(1+p)

X (i): quantization target vector

i: number of vector elements

S _t (i)、S _t+1 (i) The method comprises the following steps Parameter vector

t: time (frame number)

p: weighting factor (fixed)

d (i): the previous frame decodes the vector.

16. The CELP-type speech signal encoder of claim 14, further comprising:

a unit for decoding the current encoding target frame to generate a decoding vector;

a second distortion calculator for calculating coding distortion based on the decoding vector and the parameter vector of the coding object frame;

and a vector smoothing unit configured to smooth the parameter vector of the current encoding target frame supplied to the target extraction unit when the encoding distortion is smaller than a reference value.

17. The CELP-type speech signal encoder of claim 16, wherein the second distortion calculator calculates the perceptually weighted coding distortion according to

Ew＝∑{(V(i)-S _t (i)) ² +p{V(i)-(d(i)+S _t+1 (i))/2} ² }

Ew: audioically weighted coding distortion

S _t (i)，S _t+1 (i) The method comprises the following steps Input vector

t: time (frame number)

i: number of vector elements

V (i): decoding vectors

p: weighting factor

d (i): the vector is decoded from the previous frame.

18. The CELP-type speech signal encoder of claim 14, wherein the vector quantization unit comprises:

a plurality of codebooks correspondingly arranged to each stage of the multi-stage vector quantization and storing a plurality of code vectors;

a unit for calculating the distance between the quantization target vector or the prediction error vector thereof and the code vector stored in the level 1 codebook to obtain the level 1 code;

an amplitude storage unit for storing an amplitude expressed by a scalar corresponding to a code vector stored in the level 1 codebook;

a multiplication unit for taking out the amplitude corresponding to the 1 st level coding from the amplitude storage unit before the 2 nd level coding, and multiplying the amplitude by the code vector stored in the 2 nd level codebook; and

a difference between a decoded vector obtained by decoding the 1 st-order code and a code vector stored in the 2 nd-order codebook and obtained by multiplying the amplitude is calculated, and a unit of the 2 nd-order code is obtained.

19. The CELP-type speech signal encoder of claim 4, wherein the distortion calculator comprises:

means for calculating the power of a signal obtained by synthesizing the adaptive code vector in the synthesis filter and a filter coefficient autocorrelation matrix forming the synthesis filter, and calculating a 1 st matrix obtained by multiplying each element in the autocorrelation matrix by the power,

means for synthesizing signals obtained by synthesizing the adaptive vectors in the synthesis filter in reverse order of time, calculating a 2 nd matrix by taking a vector product of the synthesized signals in reverse order of time, and

differentiating the 1 st matrix from the 2 nd matrix to generate a 3 rd matrix to compute a unit of distortion.

20. A communication device comprising a CELP-type speech signal encoder of claim 4.

21. A CELP-type speech signal decoder, the CELP-type speech signal encoder comprising:

diffusion vector generating means for the excitation vector generating means;

a noise codebook for selecting a diffusion mode according to a noise code number determining a diffusion mode combination number and a pulse vector combination number and generating a pulse vector;

a synthesis filter for generating a synthesized speech using the excitation vector outputted from the excitation vector generation device as a noise code vector; and

an adaptive codebook storing adaptive codevectors;

the diffusion vector generation device includes:

a pulse vector generating unit that generates a pulse vector having a pulse with a polarity unit on a certain element of a vector axis;

a diffusion pattern storage unit storing a plurality of diffusion patterns;

a selection unit that selects one diffusion pattern from the plurality of diffusion patterns stored in the diffusion pattern storage unit; and

and a pulse vector diffusion unit which generates a diffusion vector by performing convolution operation of the selected diffusion pattern and the pulse vector.

22. The CELP-type speech signal decoder according to claim 21, wherein a diffusion pattern obtained by pre-learning and having less quantization distortion generated when quantizing the noise source information vector is stored in a storage unit in the excitation vector generation apparatus.

23. The CELP-type speech signal decoder of claim 22, wherein the storage unit in the excitation vector generation means stores at least one diffusion pattern derived by learning for each channel.

24. The CELP-type speech signal decoder of claim 21, wherein at least one of the diffusion patterns stored in the memory unit in the excitation vector generation means in each channel is a random pattern.

25. The CELP-type speech signal decoder according to claim 21, wherein at least one of the diffusion patterns stored in the storage unit in the excitation vector generation apparatus in each channel is a diffusion pattern obtained by learning in advance and having less quantization distortion generated when quantizing a noise source information vector, and at least one of the diffusion patterns is a random pattern.

26. A communication device comprising the CELP-type voice signal decoder of claim 21.

27. A diffusion vector generation method used in a sound source vector generation device for a CELP type speech signal encoding device/speech signal decoding device, comprising the steps of:

providing a pulse vector having pulses with a polarity unit on an element of a vector axis;

providing a diffusion mode selected from a plurality of diffusion modes;

and performing convolution operation of the selected diffusion mode and the provided pulse vector to generate a diffusion vector.

28. The diffusion vector generation method of claim 27,

the step of providing the pulse vector may provide the pulse vector to the N channels, respectively; and is

The step of providing a diffusion pattern may provide a plurality of diffusion patterns per the N channels.

Images