JPH04299400A - Voice detector - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed description of the invention]

0001

[Industrial Application Field] In order to reduce power consumption during transmission, portable radios, etc. transmit VOX (Voice) only when there is voice and interrupt transmission when there is no voice.
Operate Switch Exchange
e) control is used, which can reduce the average power consumption during transmission by about 15%. In order to perform such a VOX function, an audio detector that detects the presence or absence of an audio signal is required before the transmission output circuit. The present invention relates to a speech detector used in such a speech encoding method.

0002

2. Description of the Related Art A description will be given on the premise that a voice detector is applied to VOX control of a digital cordless telephone device. This digital cordless telephone device uses 32 kb/s adaptive differential pulse coding (ADPCM) as a voice encoding method (CODEC).
erential Pulse Code Mo
duration) is used. Furthermore, the processing delay time in this device is required to be 7 msec or less. FIG. 6 is a block diagram of a conventional audio detector.
Input audio signal a 8-bit quantized with z sampling
This is a voice detector that divides the signal into 20 msec frame units (160 samples), determines the presence or absence of voice, and outputs a voice/silence flag. Audio input signal a is sent to DC component suppressor 1
The DC component is removed by the broadband filter No. 1, resulting in a signal b, which is applied to each of the following circuits.

[0003] In the high level power detector 12, 20 ms
Divide the audio interval of ec into 5 subframes (32 samples) every 4 msec, and perform the following (1) for each subframe.
Short-term power Psk is calculated using the formula. However, xi is the filter output and k is the subframe number. Power detection is performed for the calculated Psk of each subframe using the power threshold Th2 (-30 dBm0) as shown in the following equation. When Psk≧Th2, D2k=1
(2) When Psk<Th2, D2k=
0 (3) Furthermore, the weighted sum D2 of equation (4) is obtained, and this is output as the signal c as the detection result of one frame.

In the low level power detector 13, a power threshold value T is determined for the short period power calculated by equation (1).
Power detection is performed using h1 (-50 dBm0) as shown in the following equation. When Psk≧Th1, D1k=1
(5) When Psk<Th1, D1k=
0 (6) Similarly, take the weighted sum D1 of the following equation and output the signal d as the detection result of one frame. Also, at the same time, calculate the value of the following formula.

[0005] The zero crossing number detector 14 counts the number of zero crossings of signal b (the number where the sign bits of two consecutive samples of the audio signal have different signs), so the following equation (9) is used for each subframe. Zsk is calculated by Zero cross threshold Th3(
24), the number of zero crosses is detected as shown in the following equation. When Zsk≧Th3, DZsk=1
(10) When Zsk<Th3 DZsk
=0 (11) Similarly, the weighted sum Dz of the following equation is taken and the signal e is output as the detection result of one frame.

In the interframe power increment comparator 15, 1
The power PTn for a frame is calculated by the following equation (13). Furthermore, a comparison is made with the interframe power PT(n-1) of the previous frame to perform the next power increment detection D4, and the result is output as a signal f. When PTn≧4PT(n-1), D4 = 1
(14) When PTn<4PT(n-1)
D4 = 0 (15)

Determiner 16
Now, these signals c, d, e, and f are inputted, and a speech/silence flag indicating the voice detection result is output according to the determination theory flow shown in FIG. In Figure 7, HOT is a hangover timer (a function that sets the next few frames to be voiced when the judgment changes from voiced to silent to prevent endings of sentences).
, and the SP flag means a sound/silence flag.

[0008]

[Problems to be Solved by the Invention] The processing of the conventional voice detector described above is executed in units of 20 msec frames, resulting in a delay time of at least 20 msec.
The condition of less than sec cannot be satisfied. Also,
Conventional speech detectors are configured independently of the speech encoder, so they have drawbacks such as a large amount of processing. An object of the present invention is to detect the presence or absence of speech in a short processing time and by suppressing the delay time to 7 msec or less by effectively utilizing prediction coefficients obtained in the processing process of a speech encoder having an adaptive prediction function. The object of the present invention is to provide a voice detector that can perform the following tasks.

[0009]

[Means for Solving the Problems] The voice detector of the present invention includes:
A means for inputting prediction coefficients obtained from an adaptive predictor in a speech encoder having an adaptive prediction function and calculating an average value of the prediction coefficients in a certain interval, and a prediction obtained in advance from the occurrence distribution of the prediction coefficients. means for determining whether it is a voiced section (a section where voice exists) or a silent section (a section where only background noise exists) by comparing the coefficient threshold value with the average value, This device is characterized in that it is configured to output a voice/silence flag indicating whether there is a voice or no voice.

0010

DESCRIPTION OF THE PREFERRED EMBODIMENTS As an example, an example in which the present invention is applied to a 32 kb/s (kilobits per second) ADPCM which is a voice encoder for a digital cordless telephone device will be shown below. Figure 3 shows an ADPCM with a voice detection function to which the present invention is applied.
1 is a block diagram of a speech encoder, and FIG. 1 is a block diagram showing an embodiment of a speech detector of the present invention. First, the ADPCM encoder shown in FIG. 3 will be explained. 21 is 64kb
This is a uniform PCM converter that converts a /s μ-law PCM input signal into a linear 13-bit PCM. 22 is a subtracter 22 that obtains a difference signal k by subtracting the prediction signal j that is the output of the adaptive predictor 23 from the output of the uniform PCM converter. This difference signal k is quantized by an adaptive quantizer 24, and ADP
32 kb/s audio data is sent to the transmission path as the output of the CM audio encoder. On the other hand, the adaptive dequantizer 26 outputs a quantized difference signal m by adaptively dequantizing the 32 kb/s audio data. The adder 25 outputs a reproduced signal n by adding the quantized difference signal m and the prediction signal j. The adaptive predictor 23 uses prediction coefficients a1,
a2 is calculated and used to generate a predicted signal j from the quantized difference signal m and the reproduced signal n. The prediction coefficient a1 calculated by the adaptive predictor 23 to generate the prediction signal j,
a2 is a coefficient for predicting a sample value at a certain point in time using two adjacent past sample values, and its value is different for speech signals with large autocorrelation and background noise with small autocorrelation. This is the distribution of occurrence. This prediction coefficient a1
, a2 are input to the audio detector 27 of the present invention.

[0011] To prove this, the prediction coefficients a1, a
Examples of measuring the occurrence distribution of 2 are shown in FIGS. 4(A), (B) and 5(C), (D). In the figure, FIG. 4 (A) shows the audio signal (male voice), (B) shows the audio signal (female voice),
Figure 5(C) shows white noise, and (D) shows colored noise (-6dB/
oct). In these figures, each sample point 〇,●
The range of the prediction coefficients a1 and a2 indicated by , ◎ is larger than -0.05 and smaller than +0.05, with the sample point as the origin. Further, a sample point indicating the maximum frequency of occurrence is indicated by a mark ◎, and a sample point which takes a value of 0.1 or more when normalized by the maximum frequency of occurrence is indicated by a mark. From the results shown in FIGS. 4 and 5, it can be seen that if appropriate threshold values are given to the prediction coefficients a1 and a2, it is possible to determine whether there is a voice area or a background noise area (silence area). According to the occurrence distribution diagrams of the prediction coefficients a1 and a2 shown in FIGS. 4 and 5, the speech detector 27 determines that it is a background noise section when they have values in the range of ■ to ■ shown below, and in other cases is determined to be a sound section, and outputs voice detection flags indicated by L level and H level, respectively. ■ (0.70≦a1≦1.00) and (
-0.45<a2≦-0.35) ■ (0.
75≦a1≦1.10) and (-0.55<a2
≦-0.45) ■ (0.85≦a1 ≦1
．． 20) And (-0.65<a2≦-0.55)
■ (0.95≦a1≦1.20) and
(-0.70<a2 ≦-0.65) ■ (a
1 ≦0.75) and (a2 ≦0)

FIG. 1 is a block diagram showing an example of the configuration of a voice detector according to the present invention. The processing contents of each block in FIG. 1 will be explained. The prediction coefficients a1 and a2 are inputted to framers 31 and 32, respectively, and are framed at 5 msec intervals and provided to average value calculators 33 and 34, respectively. The average value calculators 33 and 34 calculate the average value for one frame and input it to the voice/silence determiner 35. Sound/silence determiner 3
5, if the average value of the prediction coefficients a1 and a2 falls within the range of the above thresholds ■ to ■, the voice detection flag u is set to silence (
In other cases, it is set to sound (H). The hangover processing device 3
6, a hangover process of 100 msec is performed to obtain the final voice detection output v. FIG. 2 is a time chart showing the results of checking the operation of voice detection by computer simulation. The input signal contains colored noise (-6dB/
oct) is used. In the figure, (A) shows the input signal, and (B) shows the sound/silence determination result after hangover processing. From these results, it can be seen that the present method produces good results with few malfunctions due to ambient noise. In addition, prediction coefficients a1 and (D) are shown in (C) and (D), respectively.
It shows the temporal change of a2. From these, it can be confirmed that the values of the prediction coefficients a1 and a2 are different between the voice section and the background noise section.

[0013]

[Effects of the Invention] As explained in detail above, by implementing the present invention, the processing time required for voice detection processing can be reduced to about 50%.
msec, and because it efficiently uses the coefficients obtained in the ADPCM processing process, it can be realized with small-scale hardware (processing amount is 15% of ADPCM), so it has an extremely large practical effect. be.

[Brief explanation of drawings]

FIG. 1 is a block diagram of a speech detector of the present invention.

FIG. 2 is a time chart showing the operation of the present invention.

FIG. 3 is a block diagram of an ADPCM encoder with added speech detector of the present invention.

FIG. 4 is an occurrence distribution diagram of prediction coefficients a1 and a2.

FIG. 5 is an occurrence distribution diagram of prediction coefficients a1 and a2.

FIG. 6 is a block diagram of a conventional voice detector.

FIG. 7 is a conventional decision logic flowchart.

[Explanation of symbols]

11 DC component suppressor 12 High level power detector 13. Low level power detector 14 Zero crossing number detector 15 Interframe power increment comparator 16 Judgment device 21 Uniform PCM converter 22 Subtractor 23 Adaptive predictor 24 Adaptive quantizer 25 Adder 26 Adaptive inverse quantizer 27 Audio detector 31, 32 Framerizer 33, 34 Average value calculator 35 Sound/silence detector 36 Hangover processing device

Claims (1)

[Claims] 1. Means for inputting prediction coefficients obtained from an adaptive predictor in a speech encoder having an adaptive prediction function and calculating an average value for each predetermined interval for the prediction coefficients, and generating the prediction coefficients. means for determining whether the section is a sound section or a silent section by comparing a prediction coefficient threshold value obtained in advance from a distribution with the average value; A voice detector configured to output a voice/silence flag indicating whether the voice is active or silent.