JP2558876B2 - Speech coding device - Google Patents

Description

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice encoding device used for digital communication, voice mail and the like.

2. Description of the Related Art A conventional speech coding apparatus, as shown in FIG. 3 (a), includes a predictor 1, a quantizer 2, and a multiplexer 3 for multiplexing transmission data from the predictor 1 and the quantizer 2. Encoder 14 with
Then, as shown in FIG. 3 (b), a synthesis for synthesizing the transmission data from the decomposing unit 4, the quantity particle demultiplexing unit 5, the demultiplexing unit 5 and the dequantizing unit 5 for decomposing the transmission data from the encoder 14. Bowl 6
And a compound device 15 having

As shown in FIG. 4, the predictor 1 has a short-term predictive analyzer (linear predictor) 8 in order to remove a correlation between adjacent sample values of an analog / digital (A / D) converted voice signal. The short-time prediction filter coefficient is obtained by the following, and the short-time prediction error is obtained by the short-time prediction filter 7.

Then, in order to remove the periodic correlation of the sound source pitch from the output of the short-term prediction filter 7, the long-term prediction analyzer (pitch predictor) 10 obtains the long-term prediction filter coefficient, and the long-term prediction filter 9 predicts it. Find the error. This prediction error is input to the quantizer 2.

On the other hand, the short-term prediction filter coefficient and the long-term prediction filter coefficient are quantized by the quantizer 11 and input to the multiplexer 3 as transmission data.

As shown in FIG. 5, the quantizer 2 includes a vector quantizer 12 that vector-quantizes an input prediction error, and a codebook 13 that prepares a plurality of constant-length signal sequences (referred to as representative vectors). Then, the square distance between the input prediction error signal sequence and each representative vector is calculated, and the number of the representative vector having the smallest value is transmitted as the quantized value for transmission data. This transmission data is multiplexed with the transmission data (coefficient of the prediction filter) from the predictor 1 by the multiplexing unit 3,
It is sent to the decoder.

On the other hand, in the decoder, the separator 4 separates the output of the quantizer 2 (representative vector number) and the output of the predictor 1 (filter coefficient). The output of the quantizer 2 is input to the inverse quantizer 5, and the output of the predictor 1 is input to the combiner 6. The inverse quantizer 5 decodes the signal sequence corresponding to the input representative vector number using a codebook to obtain a prediction error. This prediction error is calculated by the combiner 6 in the predictor 1
The audio signal is input to the filter having the inverse characteristic of that of the above, and the audio signal is decoded.

DISCLOSURE OF THE INVENTION Problems to be Solved by the Invention However, in the above-described conventional speech coding apparatus, since the same processing is always performed on any part of the speech signal regardless of the characteristics of the speech, in speech coding at a low bit rate. Has a problem that an appropriate encoding process according to the characteristics of the voice is not performed, the voice quality after decoding is not good, and the encoding efficiency cannot be improved.

The present invention solves such a conventional problem, and an object of the present invention is to provide a speech coding apparatus with improved coding efficiency.

Means for Solving the Problems In order to achieve the above object, the present invention provides a classifier with a stationary part and a transient part according to the characteristics of a short-term prediction error after short-term prediction of an input voice signal by the short-term predictor. Divided into parts,
With this classification output, either one of the long-term predictor and the constant enveloper is used to perform the processing according to the characteristics of the voice and improve the coding efficiency.

Operation According to the present invention, with the above configuration, the output of the classifier causes a waveform similar to that of a voice signal to continue, a stationary portion with a strong correlation between pitch periods, and a transient with a strong amplitude change of the signal and a weak correlation between pitch periods. Classify into divisions.

Then, in the stationary part, since the correlation between pitch periods is strong, the correlation between pitch periods is removed using long-term prediction. From this process, the long-term prediction filter coefficient and the long-term prediction error signal are output.

Also, in the transient part, the amplitude change of the signal is large,
In particular, at the rising and falling edges of the audio signal, the amplitudes at the beginning and the end of the section are greatly different, so the amplitude envelope is detected, and the amplitude envelope information is separated to form a constant envelope. From this process, the amplitude envelopment information and the fixed-envelope simple sense prediction error signal are output.

By this processing, the signal input to the quantizer has the correlation removed or the constant envelope,
By using a quantizer based on vector quantization or the like, there is an effect that efficient coding can be performed even at a low bit rate.

Embodiment FIG. 1 shows an embodiment of the present invention. In FIG. 1 (a), 21 is a short-term predictor that performs short-term prediction of a speech signal, 22 is a classifier that classifies according to the characteristics of the short-time prediction error signal output from the short-term predictor 21, 23a, 23b Is an interlocking switch that switches processing by the output of the classifier 22, 24
Is a long-term predictor that performs long-term prediction with the short-term prediction error signal from the short-term predictor 21, and 25 is a constant envelope that detects the amplitude envelope of the signal with the short-time error signal from the short-term predictor 21. A constant enveloper that performs the quantization, 26 is a quantizer that quantizes the long-term prediction error signal from the long-term predictor 24 or the amplitude envelope information from the constant envelope 25, and 27 is a multiplexer, The multiplexer 27 outputs the output signal a from the short-term predictor 21, the output signal b from the classifier 22, the output signal c from the long-term predictor 24, the output signal d from the constant enveloper 25, and the quantizer 26. The output signal e from is multiplexed and the transmission data is output. The encoder 34 is configured by the ones to which the above reference numerals 21 to 26 are attached.

On the other hand, in FIG. 1B, 28 is a classifier for separating the transmission data from the encoder 34 into signals a to e, 29 is a dequantizer for dequantizing a signal sequence from the signal e, and 30a, 30b. Is the signal b
A reference numeral 31 is a long-time synthesizer for performing a long-time synthesis from the signal c, 32 is an envelope regenerator for reproducing an amplitude envelope from the signal d, and 33 is a synthesizer for performing a short-time synthesis from the signal a. . And, the decoder 35 is constituted by the above-mentioned reference numerals 28 to 33.

The audio signal as shown in FIG. 2 (a) is A / D-converted, and the encoder 34 is arranged at constant time intervals (hereinafter referred to as frames).
When input to, the short-term predictor 21 removes the correlation between adjacent samples for each frame, and obtains the signal a indicating the prediction filter coefficient and the short-term prediction error.

The classifier 22 uses the prediction error output from the short-term predictor 21 to classify into two types, a stationary part (a) and a transient part (b), as shown in FIG. 2 (b). The interlocked switches 23a and 23b are switched according to the result of classification by the classifier 22, and either the long-term predictor 24 or the constant enveloper 25 is selected.

In the steady part (b), the output signal of the short-term predictor 21 is input to the long-term predictor 24 via the interlocking switch 23a set to the solid line position.
Remove the correlation between pitch periods. The long-term predictor 24 outputs a long-term prediction filter coefficient c and a long-term prediction error signal. The long-term prediction filter coefficient c is input to the multiplexer 27 and becomes transmission data. The long-term prediction error signal is sent to the quantizer 26 via the interlocking switch 23b set to the solid line position.
Is input to

On the other hand, in the transient section (b), the output signal of the short-time predictor 21 is input to the constant enveloper 25 via the interlocking switch 23a switched to the dotted line position, and the high-order constant enveloper 25 amplitude Envelope d is detected. Here, for example, the amplitude envelope may be approximated by a high-order polynomial, and the coefficient of this polynomial may be used as the amplitude envelope information d. This amplitude envelope information d is a multiplexer
It is input to 27 and becomes transmission data. The short-term prediction error signal is constant-enveloped and input to the quantizer 26 via the interlocking switch 23b set at the dotted line position.

The quantizer 26 includes a long-term predictor 24 or a constant enveloper.
The input signal from 25 is quantized using, for example, vector quantization. This input signal has a correlation removed or a constant envelope, so that it is a signal suitable for vector quantization, and efficient quantization can be performed.

On the other hand, in the decoder 35, the transmission data (a to e) transmitted from the encoder 34 is separated by the separator 28, and the signal e dequantized by the dequantizer 29 is combined for a long time. The short-term prediction error signal is encoded by the device 31 or the envelope regenerator 32, and the audio signal is encoded by each frame by the short-time synthesizer 33. As described above, according to the above-described embodiment, since the coding process can be changed by the classification output of the classifier 22, there is an advantage that the coding efficiency can be improved.

EFFECTS OF THE INVENTION As described above, according to the present invention, the characteristics of the voice are classified into the stationary part and the transient part by utilizing the characteristics of the voice, and the encoding process is changed by the classification output. According to this, there is an effect that an appropriate encoding process can be performed and the encoding efficiency can be improved.

[Brief description of drawings]

FIG. 1 is a schematic block diagram of a speech coding apparatus according to an embodiment of the present invention, FIG. 2 is an explanatory diagram of input signals and output signals of speech signals, and FIG. 3 is a schematic block diagram of a conventional speech coding apparatus. FIG. 4 is a schematic block diagram of a conventional predictor, and FIG. 5 is a schematic block diagram of a conventional quantizer. 21 ... Short-time predictor, 22 ... Classifier, 23a, 23b ... Changeover switch, 24 ... Long-term predictor, 25 ... Constant enveloper,
26 ... Quantizer, 27 ... Multiplexer.

Claims (1)

(57) [Claims] 1. A short-time predictor for performing a short-time prediction of a speech signal to remove a correlation between adjacent sample values, and a short-time prediction error signal output from the short-time predictor to a stationary part and a transient part. A classifier for classifying, a long-time predictor that removes the correlation between pitch periods, a constant envelop that detects an amplitude envelope and constant-envelopes, and a quantum output of the long-term predictor or the constant-envelope And a quantizer for converting the output of the short-term predictor to the long-term predictor or the constant envelope quantizer according to the output of the classifier, and the output is quantized by the quantizer Encoding device.