JPS63237100A - 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 (Field of Industrial Application) The present invention relates to a voice detector, and particularly to a voice detector suitable for application to a digital voice insertion system or a voice packet system in the field of digital communications.

(Description of Prior Art) Conventionally, this type of voice detection technology is described in, for example, Document I:
``1975 National Institute of Electronics and Communication Engineers General Conference J, 175
3 (a method of voice detection using zero-crossing frequency), Arakai
Taku and Ochiai, Airi.

First, prior to explaining the voice detector of the present invention, this conventionally known voice detector will be explained.

FIG. 2 is a block diagram showing the voice detector disclosed in this document (2), which is configured to detect voice for each -sample. The input sample signal received at the input terminal 21 of this voice detector is input to an amplitude detection section 22 for signals having relatively large amplitudes such as vowels. On the other hand, the offset of this input sample signal is removed by a direct current (DC) suppression circuit 23, and then an arbitrary constant value a is applied to the input sample signal by an adder 24.
is added and only its sign bit is extracted and input to the zero-crossing detection unit 25 for fricative consonants.

In this amplitude detection section 22, first, a comparator circuit 26 causes a counter 27 to count up or count down depending on whether the absolute value of the input sample signal exceeds a certain value (e), and this count value tn value circuit 28 increments a certain value (e). A comparison is made with a threshold value THv, and when the threshold value THv is exceeded, an output αV having a logical value of “1” is generated.

On the other hand, in the zero crossing detection section 25, the sent code bit is sent to a delay circuit 29 and an OR circuit (08 circuit) 30.
Both circuits 29 and 30 take -hv with the sign bit before the sample. Depending on whether the results are inconsistent, the counter 31 is counted down or counted up to count the number of times the input crosses (-a), and the count value is set to a predetermined threshold in the threshold circuit 32 to determine whether or not it exceeds a certain value. Check by comparing with THz and find the threshold @ T
If the frequency exceeds Hz, the configuration is such that an output α2 of logical value “1” is generated.

Then, the logical sum α of these two outputs αV and α2 is outputted from the logical sum circuit 33, this output α is supplied to the hangover control circuit 34, the hangover time is added, and the output αout of the audio detector is output from the output terminal 35. is configured to print. Note that the period in which the output αout is “1” is a sound period, and the period in which the output αout is “rQJ” is a silent period.

The hangover control used in this conventional voice detection technology means that the signal α generated in the voice/no-speech determination section located before the hangover control section 34 changes from "1" to "0".
Even if , the output from the detector is controlled so that it continues to be output as rlJ for a certain period of time. In this way, the purpose of maintaining the output of "1" for a certain period of time is to add this hangover time to the original duration of the output "1" so that the audio part that cannot be detected by the audio detector is The purpose is to prevent omissions.

This hangover time conventionally reached a time period of about 150 to 200 m5.

Therefore, if the detection sensitivity of the voice detector can be made high in advance, this hangover time can be shortened, and the voice detection process can be performed accurately and in a short time.

Incidentally, especially in digital voice insertion systems and voice packet systems, there is a demand for voice detectors that can shorten hangover time without deteriorating voice quality.

In this type of system, as is well known from various experimental results, which need not be exemplified, as the hangover time becomes longer,
In general, the length of time that the voice detector continues to determine that there is a sound, that is, the burst length, tends to become longer. In other words, the proportion of the occupied portion of the total call time (this is called the activity of the voice detector) increases. This increase in activity is not very desirable for the system described above, which aims to make effective use of the transmission path by transmitting other voices and data during silent parts. What is desired is a voice detector that can be kept to a minimum and has low activity. For this reason, it is desirable to increase the sensitivity of the voice detector and, as a result, reduce the hangover time.

(Problem to be Solved by the Invention) However, the zero-crossing method used in the conventional voice detection technology described above only counts the number of times the input sample crosses a certain value during the sample period, so it is not accurate. The spectral distribution of the input sample sequence cannot be estimated. For example, even if line noise and a voice signal have the same number of zero crossings, the spectral distributions do not always fall flat, and even if there is an auditory significant difference in the spectral distributions, the number of zero crossings will The values may appear to be the same.

For these reasons, conventional voice detection techniques have not been able to provide a technically satisfactory result in cutting out voice sections. Furthermore, the conventional voice detection technique using the number of zero crossings has a problem in that it is only possible to significantly shorten the hangover time.

The purpose of this invention is to eliminate the problem of missing audio parts caused by the limitations of the conventional spectral distribution analysis using the number of zero crossings mentioned above, and the problem that the closing time during hangover cannot be significantly shortened even if the number of zero crossings is used. Another object of the present invention is to provide a voice detector that has excellent ability to discriminate between line noise and voice, and can cut out appropriate voice sections even during short hangover times.

(Means for Solving the Problems) In order to achieve this objective, the speech detector of the present invention provides linear processing based on autocorrelation coefficients for each predetermined frame for the received input sample signal. A linear prediction analysis unit that performs predictive analysis and outputs the average power in this frame, this autocorrelation coefficient, and a linear prediction coefficient; A linear prediction coefficient path M calculation unit that calculates and outputs a linear prediction count distance, and a first comparison that compares the above-mentioned average power with a line noise determination threshold and a voice/no-speech determination threshold and outputs a first comparison result. a second comparison section that compares the linear prediction coefficient path M with a white noise determination threshold and outputs a second comparison result; The present invention is characterized in that it includes a determination section that performs determination.

(Function) According to the configuration of the present invention as described above, since the linear predictive counting distance (referred to as LPC distance) is used, malfunctions due to line noise can be avoided, and moreover, it is highly sensitive to voice signals. Detection becomes possible. Therefore, hangover time can be shortened without deteriorating voice quality.

(Embodiments) Hereinafter, embodiments of the voice detector of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram showing an embodiment of the voice detector of the present invention. 10 is an input terminal of this voice detector, and 11 is an output terminal. The input sample signal x(i) received at the input terminal 10 (where i indicates the number of the sample in the LPG analysis frame, which will be described later) is processed by a linear predictive analysis section (hereinafter simply referred to as the LPG analysis section). .)1
Supply to 2. This LPG analysis unit 12 performs linear prediction (LPC) analysis on the input sample signal based on the autocorrelation coefficient for each predetermined frame, and calculates the average power within the frame, the autocorrelation coefficient, and the linear The configuration is configured to output prediction coefficients.

The procedure for LPG analysis will be explained with reference to the operational flowchart of FIG. In addition, in the following description and drawings, a processing step is indicated by S.

First, after starting the LPG analysis, the input sample signal x(i)
(Sl) and divides it into LPG analysis frames with blocks of a predetermined frame length, for example, several tens of samples (
S2).

Next, for each divided frame, a total of (p+1.
) self-phase opening coefficients (R,, R+, R2... days,)
(S3), where the subscript lowercase number is LP
It represents the order of G analysis; for example, R3 is called a third-order autocorrelation coefficient. This collection of autocorrelation coefficients is representatively shown as an autocorrelation coefficient matrix F, where the 0th order autocorrelation coefficient R0 is the average power of the input signal x (1) in each frame; is output separately for post-processing.

Next, using this self-phase opening coefficient, a total of (p+1) linear prediction (LPG) coefficients (1, α,) are calculated.

α2. .

Next, it is checked whether (p+1) autocorrelation coefficients and LPG coefficients have been found (S5). If it has not been determined, the process from step S4 is repeated.

The autocorrelation coefficient and PC coefficient vector σ thus obtained are output to the LPG distance calculation stage 14 of the next stage linear prediction coefficient distance calculation unit (hereinafter simply referred to as the LPG distance calculation unit) 13. Along with the average power mortar. The first comparison part 1
5 (S6).

This LPG distance calculation unit 13 is a part that performs processing corresponding to the number of zero crossings, and in this LPG distance calculation unit 13, the supplied LPG coefficient vector α and autocorrelation coefficient matrix date, as well as the data stored in a ROM etc. Standard hectory memo 1,
Standard LF) C coefficient vector β stored in 116 and prepared
The LPG distance is calculated from and output to the second comparison section 16.

This LPG distance can be found in the document: II r Audio Digital Signal Processing (Part 2), published by Corona Publishing, April 15, 1982.
As can be understood from the description in Japanese, pages 269 to 271, it is given by the following formula.

However, cr=(1,α1.α2....αF)8=(1,β
1. β2. ... β, )8 = (Japan IJ), Rii =
R1l-Jl, i, j=o~p day r = (1/N)Σx (n)-x (n+1)7
1+1, N is the number of samples in the analysis frame, R+ is LPG
The i-th autocorrelation coefficient of the analysis frame, α1 is the i-th LPG coefficient, β1 is the i-th standard LPG coefficient, and t means the centrifugation matrix.

As already mentioned, this standard LPG coefficient 81 is prepared in advance. In this example, since line noise is generally considered to have properties similar to white noise, line noise is considered to be white noise, and the linear prediction coefficients for each frame of white noise are actually measured and the average value thereof is calculated. and convert this into standard LPG
Coefficient 81. . , this 6 is just an average value (average vector), so the linear prediction coefficient of each frame of white noise is not necessarily equal to this β, so how similar is the linear prediction coefficient of each frame to 6? The LPC distance MD of equation (1) is determined as a guideline for determining whether the vehicle is present.

The procedure for calculating this LPG distance may be, for example, by calculating necessary terms in the denominator and numerator, then performing multiplication, and then performing division, but other calculation procedures may be used, and the procedure does not matter.

The first comparison section 15 is a section that performs processing corresponding to amplitude detection, and in this first comparison section 15, an average power mill is used. is compared with a preset line noise determination threshold P th+ and a voice/silence determination threshold P th2, and a first comparison result vp is output. Generally, the average level of line noise is -4
It is said to have a Jawo of 0dBmO to -50dBmO. In this example, the threshold p th is −40 dBm
It is set to an appropriate value larger than the value of O to prevent the voice detector from malfunctioning due to line noise of -40 dBmO.

Regarding the voice/silence determination threshold P th2, it is set to an arbitrary and suitable value such that most of the audio signals have a power equal to or greater than this threshold P th2 and signals below this can be judged as silence. Set. Preferably, this threshold [P th2 is set to about -50 dBmO.

Therefore, from this first comparison unit 15, the average power mill inputted with respect to the thresholds 1 pthl and P th2 set in this way. Or, when there is a relationship as shown in the following equation (2), the corresponding output signals are used as the first comparison result Vp! It is configured to output.

On the other hand, the second comparing section 17 compares the LPC distance with a predetermined white noise determination threshold oth and outputs a second comparison result Vdt. In this embodiment, it is preferable to calculate the LPC distance of each frame of white noise, check to what extent the value reaches, and calculate the probability that each LPG distance (a D exceeds a certain value A is 0.01%). It is best to find in advance a value for the relevant value such that , and use this as the white judgment threshold Dth. This threshold value Dth changes depending on what percentage is set.

In this embodiment, based on the difference in size between this LPG distance and the threshold value oth, a comparison result d of logical values determined by the following equation (3) is output as a result of comprehensive judgment.

Both the first comparison result Vp and the second comparison result Vd obtained as described above are sent to the sound/non-sound determining section 18, and there, the sound/non-sound determination unit 18 determines whether the sound is present or not based on the first and second comparison results Vp and Vd. Performs a comprehensive judgment of silence and generates a sound/no-sound judgment signal V
is generated and output to the hangover adding section 19 at the next stage. This determination is made according to the following equation (4).

In equation (4), when the judgment signal V is the logical value "]", it means there is a sound, and when it is l''OJ, it means no sound.

This is the average power mill mentioned above. And L If it is present and p th2 or more, the judgment signal V is V
=Vd, and it can be understood that whether there is a sound or not is determined only by the determination based on the LPG distance jlD.

Roughly speaking, when the average power R8 is smaller than the threshold value Pth2, the determination signal V is unconditionally set to V=◯ and becomes silent.

After adding a hangover time to the determination signal determined and output in this manner in the hangover adding section 19 as in the conventional case, the output terminal 11 outputs the output signal Vout.
However, according to the voice detection device of the present invention, in the power level region where line noise and voice coexist, the difference in the spectral envelope is used as a technique for distinguishing between line noise and voice (as described above). 1) Chu LP shown in equation
Since the technique of calculating the C distance is used, it is possible to realize a highly sensitive voice detection device i% that has a high ability to distinguish between line noise and voice signals. Therefore, hangover time can be shortened.

It is clear that the invention is not limited only to the embodiments described above, but can be subjected to many variations and modifications.

For example, it is sufficient that the first and second comparing sections store their respective threshold values in an appropriate memory, read them from the memory when an input signal arrives, and perform comparison processing in each comparing section. Its specific configuration can be easily configured according to the design using conventional Takaji technology. Also, regarding the sound/non-sound determination section, two input signals vp
As long as the configuration is such that the sum of and Vd is calculated and the result is compared with the O level, the specific configuration can be any suitable configuration depending on the design.

(Effects of the Invention) As is clear from the above description, the voice detector of the present invention has a structure in which, after LPG analysis is performed, line noise and voice signals are discriminated using the LPG distance. This reduces malfunctions caused by line noise, and
It is possible to detect audio signals with higher sensitivity than before. Due to this high sensitivity, the hangover time can be reduced to about 1/3 to 1/2 of the conventional one.

[Brief explanation of the drawing]

FIG. 1 is a block diagram for explaining the configuration of an embodiment of the voice detector of the present invention, FIG. 2 is a block diagram for explaining a conventional voice detector, and FIG. It is a flow chart of the operation provided for explanation of the operation of the voice detector of the invention. 10...Input terminal, N-...Output terminal 12.
...LPG analysis section, 13...LPG distance calculation section 1
4... LPC distance calculation stage, 15... First comparison section 16
. . . Standard vector memory + 7- . . . Second comparing section 18 . . . Sound/silence determining section 19 . . . Hangover adding section. Patent Applicant Oki Electric Industry Co., Ltd. - LPC Minute FRLP Collaborative Danger Diagram Figure 3

Claims (1)

[Claims] (1) Perform linear prediction analysis based on the autocorrelation coefficient for each predetermined frame on the received input sample signal, and calculate the average power in the frame, the autocorrelation coefficient, and the linear prediction coefficient. a linear prediction analysis unit that outputs a linear prediction coefficient distance, a linear prediction coefficient distance calculation unit that calculates and outputs a linear prediction coefficient distance from the linear prediction coefficient, autocorrelation coefficient, and a standard vector prepared in advance; a first comparison unit that compares the linear prediction coefficient distance with a determination threshold and a voice/no-speech determination threshold and outputs a first comparison result; and a second comparison unit that compares the linear prediction coefficient distance with a white noise determination threshold and outputs a second comparison result. A voice detector comprising: a comparison section; and a determination section that performs a comprehensive determination of whether there is a sound or no sound based on the first and second comparison results.