JPS63281200A - Voice section detecting system - 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 (Field of Industrial Application) The present invention relates to a method for detecting a voice section in a digital voice insertion system or a voice packet system in the field of digital communications.

(Prior Art) Conventionally, the one shown in FIG. 2 is known as a speech detector applied to the speech section detection method. According to Figure 2,
Among the input sample signals A input to the input terminal 1, a signal AI having a relatively large amplitude such as a vowel is input to the amplitude detection section 2, and a signal A2 due to a fricative consonant is subjected to offset removal by a DC suppression circuit 4. After that, a constant value a is added and the sign bit is extracted and input to the zero crossing detection section 3. In the amplitude detection section 2, a comparator circuit 2a compares the absolute value of the signal Al with a predetermined value θ, increases or decreases a counter 2b according to the result, and when the counter value of the counter 2b becomes larger than the threshold value THv, the threshold value The output αV from the circuit 2C is outputted to the OR circuit 5 at a high level "1". On the other hand, the zero crossing detection section 3 determines whether the sign bit inputted by the OR circuit 3a matches the sign bit one sample before, and
The counter 3b is increased or decreased depending on whether the result matches or does not match. This is equivalent to counting the number of times the input crosses (-a), and when the counter value of the counter 3b becomes larger than the threshold THz, the output α2 from the threshold circuit 2C is sent to the OR circuit 5 at a high level "1". Output. The logical sum α of the outputs αV and α2 from the threshold circuits 2C and 3C is input from the OR circuit 5 to the hangover control circuit 6, and the hangover control circuit 6 changes the output α of the OR circuit 5 from high level "1" to low level. Even if the level becomes "0", a hangover time is added in which the high level "1" is continued to be output for a certain period of time, and the output is output from the output terminal 7 as the output αout.

While this αout is at a high level “1”, there is a sound,
If the low level is "0", there is no sound (1975 General National Conference of the Institute of Electronics and Communication Engineers, 1753 "A method of sound detection using r zero crossing frequency" by Takashi Araseki and Kazuo Ochiai).

(Problem to be Solved by the Invention) However, according to the above configuration, the number of zero crossings is counted as the number of times the input sample signal A crosses a certain value (-a), so the accurate input sample sequence is also It is not an estimation of the distribution. In other words, even if the line noise and the voice signal have the same number of zero crossings, there is a possibility that there will be a significant difference in the spectral distribution perceptually, and there will be a point where the extraction of the voice section will become inaccurate. Ta. Furthermore, since zero-crossing information is used to detect voice sections, it is not possible to significantly reduce hangover time, and as a result, the proportion of voiced sections in the total call time increases, making it difficult to use digital voice insertion systems or In a system such as a voice packet system that attempts to utilize the transmission path effectively by transmitting other voice or data during silent portions, there is a problem in that further efficiency cannot be expected.

SUMMARY OF THE INVENTION In view of the above-mentioned problems, an object of the present invention is to provide a voice section detection method that can shorten hangover time and accurately detect voice sections.

(Means for Solving the Problems) In order to achieve the above object, the present invention distinguishes the line noise of the input sample signal from the voice signal, and detects the voice signal by detecting the line noise of the input sample signal. an analysis means that divides the signal into predetermined frame lengths, performs LPC analysis on each frame, and calculates the average para and LPC coefficient of each frame; and the LPC distance between the LPC coefficient and a preset standard vector and a preset threshold value. Compare with L
a first determination means that determines that there is a sound if the PC distance is greater than the threshold; and a first determination means that compares the average power with a preset threshold, and determines that there is a sound if the average power is greater than the threshold; and a third determining means that compares the average variation with an adaptive power threshold that can be updated within a predetermined range, and determines that there is a sound if the average power is larger than the adaptive power threshold.
determining means; and the first determining means. a fourth determining means that determines whether the audio signal is voiced or silent based on the determination results of the second and third determining means, and counts the number of frames of voiced and silent frames; updating means for calculating and updating an adaptive power threshold based on an instruction from the means; The adaptive power threshold is updated when consecutive frames are determined to be silent by the determining means and are determined to be sound by the third determining means and the number reaches the specified number. shall be.

(Function) According to the present invention, the analysis means divides the input sample signal into frames of a predetermined length, performs LPC analysis on the frames to calculate the average variation and the LPC coefficient, and the first determination means divides the input sample signal into frames of a predetermined length. A comparison is made between the LPC distance between the coefficient and a preset standard vector and a preset threshold value to determine whether there is a sound or no sound based on the LPC distance. Second
The third determination means compares the average power of the frame with a preset threshold value, and determines whether there is a sound or no voice based on the average power of the frame. A comparison is made with the adaptive power threshold from the updating means, and a determination is made as to whether there is a sound or not based on the adaptive power threshold. Here, the fourth determination means is the first determination means described above. Second
The determination results of the third determining means are integrated to determine whether the audio signal is voiced or not, and the voice is detected.

The fourth determining means counts the number of frames with sound and without sound, and determines whether the number of consecutive silent frames reaches a preset specified number or whether the first and second determining means determine that there is no sound. and when the number of consecutive frames determined to be sound by the third determining means reaches the specified number, the updating means is instructed to update the adaptive power threshold, whereby the updating means: The adaptive power threshold is calculated and updated.

(Embodiment) FIG. 1 is a block diagram showing an embodiment of a speech detector to which the speech section detection method according to the present invention is applied. In the figure,
1 is an input terminal, 7 is an output terminal, 10 is an LPC analysis section, 1
1 is an LPC distance calculation unit, 12 is a standard vector memory, 1
3, 14.16 is a comparison unit, 15 is an adaptive power threshold update unit,
Reference numeral 17 represents a sound/non-speech determination unit, and 18 represents a hangover time addition unit.

The LPC analysis unit 10 inputs the input sample signal A, divides it into predetermined frame lengths, performs LPC analysis, calculates the autocorrelation coefficient R1 and LPC coefficient ki of the LPC analysis frame, and calculates the autocorrelation coefficient matrix IR and the LPC coefficient. The vector 1 is sent to the LPC distance calculation unit 11, and the 0th order autocorrelation coefficient RO is set as the average power of the LPC analysis frame to the comparison unit 14.16 and the adaptive power threshold updating unit 1.
Send to 5.

The LPC distance calculation unit 11 calculates the LPC distance from the autocorrelation coefficient matrix IR from the LPC analysis unit 10, the LPC coefficient vector 1k, and the standard LPC coefficient vector 1b stored in the standard vector memory 12 according to the following formula (1). is calculated and the result is sent to the comparison section 13.

(lk-1b)・IRI・(lk-1b ”) t
D-(1) lk ・IR-Ikt However, lk-(1, kl, 2, ..., kp
)lb-(1,bl,b2,...,bp)IR-
(Rlj) Rlj−R11−j l (t, j−0+
..., p) Σ x (n) ・ x (
n+1)R1-□ Note that N is the number of samples in the LPC analysis frame, X(1) is the sample value of the input signal, p is the analysis order of LPC analysis, and R
1 is the first-order autocorrelation coefficient of the LPC analysis frame, kl
is the first-order LPC coefficient, bl is the first-order standard LPC coefficient, n is the number of the LPC analysis frame, and t is a symbol indicating vector transposition. (For formula (1) above, refer to formula 9-21 of "Sound Digital Signal Processing (Part 2) JP269-P271," published by Corona Publishing, April 15, 1980.
Author: R, Rablner, R, W, 5ch
af'er s Translator: Hisaki Suzuki).

The standard LPC coefficient vector 1b is stored in the standard vector memory 2, and this standard LPC coefficient vector 1b is determined from actual measurement data of white noise as line noise.

The comparison unit 13 compares the LPC distance from the LPC distance calculation unit 11 with the threshold value Dth, and according to the following formula (2),
If the LPC distance is greater than the threshold Dth, the LPC
The determination signal VD of whether there is a sound or no sound based on the distance is set to a high level "1" indicating the presence of sound, and when the LPC distance is equal to or smaller than the threshold value pth, the determination signal VD is set to a low level "0" indicating no sound. The signal is sent to the sound/silence determining section 17.

VD −1(D>Dth) (2)0 (D
≦D th) The threshold D-th has a small value if the spectral envelope shapes of the LPC coefficient vector 1k and the standard LPC coefficient vector 1b are similar in LPC distance, and has a large value if they are different. Therefore, if the LPC distance is large, the input frame can be determined as a colored signal, that is, an audio signal, rather than a white signal, and the audio signal can be detected using this difference in spectral distribution. This is provided for the LPC distance in order to

The comparison unit 14 compares the average power of the LPC analysis frame from the LPC analysis unit 10, that is, the 0th order autocorrelation coefficient RO, with the first threshold p thi and the second threshold Pth2, and calculates the following ( 3) According to formula, the autocorrelation coefficient RO is set to the threshold p
If it is larger than tht, the sound/non-sound determination signal Vp based on the average power of the LPC analysis frame is set to a high level "1" indicating sound, and the autocorrelation coefficient RO is set to the first threshold P.
equal to or smaller than thl, and second threshold P th2
If the autocorrelation coefficient RO is smaller than the threshold value Pth2, the detection signal Vp is set to a low level "-1" indicating silence. ” as the sound/non-sound determination unit 17.
Send to.

1 (RO>Pthl) Vp -0 (Pthl ≧RO≧Pth2) (
3)-1(Pth2 > RO) (However, pthl > Pth2) The adaptive power threshold updating unit 15 inputs the 0th order autocorrelation coefficient RO which is the average power of the LPC analysis frame, and also inputs the 0th order autocorrelation coefficient RO which is the average power of the LPC analysis frame, and According to the following equation (4) according to the instruction of the control signal CAP sent from the determination unit 17, the adaptive power threshold PAP is calculated and updated based on the average level of consecutive MAP LPC analysis frames according to the conditions described later,
This adaptive power threshold PAP is sent to the comparator 16. Further, the variable range of the adaptive power threshold PAP is the threshold P thl
When the value is less than or equal to P th2 and greater than the threshold P thl, the value is set to the value of the threshold P thl, and the threshold P th2 is set to the value of the threshold P thl.
When it becomes smaller, it is held at the value of threshold P th2.

l NAP P AP-M APX -Σ D R0(S) (
4) NAP S-1 However, MAP is a constant, NAP is a preset specified number of consecutive frames counted by a counter described later, and D
R0(S) indicates the average power of consecutive MAP LPC analysis frames.

The comparison unit 16 inputs the adaptive power threshold PAP and the 0th order autocorrelation coefficient RO, which is the average power of the LPC analysis frame, and compares the two, and calculates the autocorrelation coefficient RO according to the following equation (5). is larger than the adaptive power threshold PAP, the judgment signal VAP based on the adaptive power threshold is set to a high level "1" indicating a sound, and the autocorrelation coefficient RO is set to a high level "1" that indicates a sound.
If it is equal to or smaller than AP, the determination signal VAP is set to a low level "0" indicating silence, and the sound/silence determination unit 1
Send on 7.

VAP -1 (RO>PAP) (5) 0
(RO≦P AP) (However, P thl ≧PAP≧Pth2) The voice/silence determination unit 17 uses the determination signal VD based on the LPC distance HD from the comparison unit 13 and the average power of the LPC frame from the comparison unit 14. By inputting the judgment signal ■P based on the judgment signal P and the judgment signal VAP based on the adaptive power threshold PAP from the comparator 16, the judgment signals VD, VP, VAP are calculated according to the following equation (6).
If the sum is greater than 0, it is determined that there is a sound, and the sound/silence determination signal V is set to a high level "1", and the determination signal VD,
If the sum of VP and VAP is rOJ or smaller than rOJ,
It is determined that there is no sound and the sound/no-sound judgment signal V is set to a low level rOJ.
It is sent to the hangover time addition unit 18 as a hangover time adding unit 18.

V-1(VD +VP +VAP>0) (8)0
(VD +VP +VAP≦0) Furthermore, the adaptive power threshold updating unit 15 has a counter 17a for setting conditions for calculating the adaptive power threshold PAP,
LPC determined to be silent by sound/silence determination signal V
As long as analysis frames occur continuously, the number of LPC analysis frames is counted in units of NAP specified in advance by the counter 17a, and if an LPC analysis frame whose sum of judgment signals VD and VP is larger than "0" occurs, For example, at that point, the counter value NC is returned to "0". On the other hand, the judgment signal VD,
If the sum of VP is "0" or less than "0", the judgment signal VAP
If a frame with a high level "1" occurs, the counter value NC is set to "1", and as long as this state continues, the counter 17a is incremented by counting NAP pieces.

In this way, when the counter value NC of the counter 17a becomes equal to the predetermined value NAP, it is immediately determined whether the sum of the judgment signals VD, VP, and VAP (7) is "0" or "
Whether NAP consecutive LPC analysis frames smaller than "0" have occurred, or whether the sum of detection signals VD and VP is "0",
If NAP consecutive LPC analysis frames are smaller than "0" and have a high level "1" in which the detection signal VAP indicates sound, the control signal CAP is changed to the adaptive power threshold updating unit 15.
and returns the counter value NC to "0".

The hangover time adding section 18 includes the sound/silence determining section 17.
If the judgment signal V is a high level "1" indicating the presence of sound, a hangover time is added and the final judgment signal V out is output from the terminal. Output from 7.

3(a), (b), and (c) are diagrams for explaining the relationship between the average power of the LPC analysis frame, the adaptive power threshold PAP, the voice/silence determination signal V, and the operation of the counter 17a. It is an explanatory diagram. FIG. 3(a) shows the temporal change in the average power of the LPC analysis frame of input sample signal A,
It shows the time change of the adaptive power threshold PAP, where:
The X section is a voice signal section, the Y section is a line noise section, and the Y' section is a line noise section where the average power temporarily increases significantly. FIG. 3(b) shows the output result of the sound/silence determination signal V corresponding to FIG. 3(a), and FIG. 3(c) shows the output result of the determination signal VD corresponding to FIG. 3(a). , VP ,
The counter value NC of the counter 17a used to calculate VAP and adaptive power threshold PAP is shown. In addition, NAP is 6
is specified.

Here, the operation of the above configuration is shown in FIG. 4(a).

This will be explained using the flowchart in (b).

First, the LPC analysis unit 10 divides the input sample signal A input from the input terminal 1 into predetermined frame lengths (S1).
, LPC analysis is performed to calculate the autocorrelation coefficient matrix IR and LPC coefficient vector 1k of the LPC analysis frame (S
2) Sends it to the LPC distance calculation unit 11 and LP
The zero-th order autocorrelation coefficient RO indicating the average power of the C analysis frame is sent to the comparison unit 14 16 and the adaptive power threshold updating unit 15 . The LPC distance calculation unit 11 calculates the LPC distance from the autocorrelation coefficient matrix IR, the LPC coefficient vector 1k, and the standard LPC coefficient vector lb stored in the standard vector memory 12 according to equation (1), and calculates the result. is sent to the comparison unit 13 (S3). The comparator 13 does not compare the LPC distance with the threshold value Dth according to equation (2) (S4), and sets the determination signal VD to a high level "1" if the LPC distance HD is greater than the threshold value Dth. S5)
, LPC distance is equal to or smaller than the threshold value, the determination signal VD is set to a low level "0" (S6) and sent to the sound/silence determination section 17. In addition, the comparison unit 14 does not compare the magnitude of the 0th order autocorrelation coefficient RO indicating the average power of the LPC analysis frame with the first threshold value Pthl (S7), so that the autocorrelation coefficient RO indicates the average power of the LPC analysis frame. If the autocorrelation coefficient RO is smaller than the first threshold P thl, the determination signal vP is set to a high level "1" (S8), and if the autocorrelation coefficient RO is smaller than the first threshold P thl, it is set to the second threshold P th2. No comparison between S
9) If the autocorrelation coefficient RO is greater than or equal to the second threshold Pth2, the determination signal VP is set to a low level "0". 10) The autocorrelation coefficient RO is set to the second threshold Pth2.
If it is smaller than th2, the determination signal vp is set to "-1" (Sll) and sent to the voice/silence determination section 17. Furthermore,
The comparison unit 16 eliminates the magnitude comparison between the adaptive power threshold PAP (initial value is Pthl) and the autocorrelation coefficient RO according to equation (5).
If it is larger than P, the judgment signal VAP is set to high level "1".
If the autocorrelation coefficient RO is equal to or smaller than the adaptive power threshold PAP (S13), the determination signal VAP is set to a low level "0" (S14) and sent to the voice/silence determination section 17.

The sound/silence determining section 17 receives determination signals VD and VP.

Input VAP and generate judgment signals VD and VP according to equation 6).

The sum of VAP and 0 are not compared (S15), and if this sum is equal to or smaller than "0", the sound/silence determination signal V is set to a low level "0" (S16) and the counter 17a is 1 is added to the counter value NC (SL?), while the judgment signals VD and VP.

If the sum of VAP is greater than "0", the judgment signal V is set to a high level "1" 818), and the sum of judgment signals VD and vP is not compared with "0" (S 19), and this sum is is 0
If it is larger than that, set the counter value NC of the counter 17a to "
0" (S20). If the sum of the judgment signals VD and VP is equal to or smaller than rOJ, it is not possible to determine whether the previous LPC analysis frame was a sound or a soundless frame.5
21), if there is no sound, set the counter value NC to "1" 522), and if there is sound, set the counter value NG to "1".
” is added (S23). After making the above judgments,
The level of the sound/silence determination signal v of the current LPC analysis frame is stored so that the next LPC analysis frame can refer to it in S21 (S24), and the average power of each of the consecutive NAP LPC analysis frames is also stored. Remember S 2
5) It is determined whether the counter value NC and the designated value NAP match or do not match (S26).

If the counter value NG and the specified value NAP match, the adaptive power threshold updating unit 15 calculates the adaptive power difference value PAP according to equation (4) by setting the control signal CAP to a high level "1". (S27). On the other hand, counter value NC
has not reached the designated value NAP, the control signal CAP is set to a low level rOJ, and if the sound/no-sound determination signal V is a high level "1" indicating the presence of sound, the hangover time addition unit 18 sets the hangover time to The output terminal 7 outputs the final sound/non-sound determination signal V out (82g). The above operations are repeated sequentially from 81 to 528.

According to this embodiment, the power of the voice signal varies over a large range, but line noise is within a certain power range. A voice signal is detected in a region where only a voice signal is present when the average power is greater than a predetermined threshold P thl; In areas where audio and audio signals coexist, detection based on differences in spectral distribution due to LPC distance and detection using adaptive power threshold PAP are used together, reducing the influence of malfunctions due to line noise. In addition, it is possible to detect appropriate speech intervals.

(Effects of the Invention) As explained above, according to the present invention, in order to distinguish between line noise and voice signals, LP coefficients and standard vectors are
By comparing the C distance with a preset threshold, it is possible to determine whether there is sound or silence based on the difference in spectral distribution, and whether the number of consecutive silent frames reaches a preset specified number is determined based on the LPC distance. When the number of consecutive frames in which the determination based on the average power is silent and the determination based on the adaptive power threshold reaches the specified number, the adaptive power threshold is updated within a predetermined range.
Speech/silence can be determined by comparing the adaptive power threshold that correctly estimates the line noise level with the average power of the frame. Therefore, malfunctions caused by line noise can be minimized, and voice sections can be detected accurately.
This has the advantage of greatly reducing hangover time.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing an embodiment of a speech detector to which the speech interval detection method according to the present invention is applied; FIG. 2 is a block diagram showing an embodiment of a speech detector to which the conventional speech interval detection method is applied; Figures 3 (a), (b), and (c) are LPC
An explanatory diagram for explaining the relationship between frame average power, adaptive power threshold, voice/silence determination signal, and counter operation;
FIG. 4(ac, b) is a flowchart for explaining the operation according to the present invention. In the figure, 1...input terminal, 7...output terminal, 10...
・LPC analysis section, 11...LPC distance calculation section, 12.
... Standard vector memory, 13, 14. 16... Comparison unit, 15... Adaptive power threshold update unit, 17... Sound
Silence determination section, 18... Hangover time addition section.

Claims (1)

[Claims] In a voice section detection method that distinguishes between line noise and voice signals in an input sample signal and detects the voice signal, the input sample signal is divided into predetermined frame lengths, and each frame is linearly predicted (hereinafter referred to as , LPC) for calculating the average power and LPC coefficient of each frame; and LPC of the LPC coefficient and a preset standard vector.
a first determination means that compares the distance with a preset threshold and determines that there is a sound if the LPC distance is greater than the threshold; and a first determination means that compares the average power with a preset threshold and determines the average power. a second determination means that determines that there is a sound if the average power is greater than the threshold, and a second determination means that compares the average power with an adaptive power threshold that can be updated within a predetermined range, and if the average power is greater than the adaptive power threshold; a third determining means for determining whether there is a sound if the audio signal is present; a fourth determining means for counting the number of frames; and an updating means for calculating and updating an adaptive power threshold based on an instruction from the fourth determining means, the updating means configured to count the number of consecutive silent frames. reaches a specified number set in advance, or the number of consecutive frames in which the first and second determining means determine that there is no sound, and the third determining means determines that there is sound reaches the specified number. A speech interval detection method characterized in that an adaptive power threshold is updated when the threshold value is reached.