JPH01123297A - Voice detection circuit - Google Patents

Abstract

PURPOSE: To detect a voice included in an input signal on conditions different by noise levels by updating a threshold in accordance with the noise level to decide sound/silence. CONSTITUTION: A voice signal is divided by a unit time, and an average power Pi of the section is calculated, and a certain value α is added to the average power Pi to calculate a value (Pi +α). The average power Pi and a threshold Ti are compared with each other to output a discridecision result X0 showing larger one. The calculated value (Pi +α) and the threshold Ti are compared with each other to output a discrimination result X1 showing larger one. This decision result X1 is inputted, and a threshold Ti+1 of the next section is set to Ti if (Pi +α) is equal to or larger than Ti , and the threshold Ti+1 of the next section is set to (Pi +α) if (Pi +α) is smaller than Ti . The threshold Ti is set to a preliminarily determined certain value T0 at the time of the start of operation and is changed to output the decision result X0 at each time of elapse of the unit time.

Description

Detailed Description of the Invention "Field of Industrial Application" This invention relates to devices that transmit or store voices uttered by humans, and devices that are controlled by voices uttered by humans. This relates to a voice detection circuit that detects when a voice has ended.

"Prior Art" Conventionally, methods for configuring a voice detection circuit used in a device for transmitting or storing voice have been proposed. For example, it can be applied to systems that have functions such as transmitting only voiced parts to reduce transmission costs, or compressing and storing silent parts to improve the efficiency of audio storage. (For example, Yoshiyuki et al. ``Silence compression effect in audio storage services'', 1932)
6 IEICE Information System Zenten, 473) The basic principle of these voice detection circuits is that the amplitude of the input signal,
Frequency components, number of zero crossings, analysis parameters such as linear analysis, etc. are used, and determination is made based on the thickness and time change t of these. (For example, Maruta, R4, etc.
al,: ”Design and Performance+n
anceofaDs I Terminal for
Domestic Applications”.

IEEE Trans,CommunlCOM-29
, pp. 337-344 (March 1981) However, the basic parameters of both methods are based on the average power (corresponding to the amplitude envelope) of the input signal, and other parameters only supplement this.

The method of determining voice/silence using average power is that if the short-term average power (average value of the square of the amplitude of the audio signal) exceeds a predetermined threshold, then that section is considered to be a voice, and the threshold value is determined. If it is less than , it is determined that there is no sound.

Furthermore, some systems incorporate an algorithm that smooths this result over time, corrects momentary silent periods in speech, and corrects sudden noises in silent states.

(For example, Yoshikawa, Ueda, "Silent interval compression processing in voice communication systems that allow delay," IEICE Ronsho 58-Wa 319
CB-105) pp90B-915) If the voice level is sufficiently high and the noise level is small, a fixed value can be used as the threshold value used for determination. However, in devices that detect voice via telephone lines, the noise level varies considerably depending on the length of the line, transmission characteristics, etc. In order to detect the voice included in the input signal under conditions with different noise levels, it is necessary to measure the noise immediately after connecting the line, and then set the threshold value for determining voice/no-speech based on this. , there is a method for making subsequent determinations.

(For example, Hama, Ochiai, ``A method of adaptive threshold voice detection circuit'', IEICE, 1972) However, if we measure noise at a specific time, it is difficult to measure the noise at what point and for how long. However, there is a problem in that speech or sudden noise that occurs at the time of noise measurement may cause errors in subsequent judgments.Also, noise measurement is performed by anticipating the point in time when it is expected that the person will not start speaking. Doing so requires careful consideration for each application form, making system design difficult.

"Means to Solve the Problem" According to this invention, a continuously input audio signal is divided into units of time (hereinafter, the description will focus on the first divided section). The average power Pi of the section is calculated, and the value obtained by adding a constant value α to the average power Pi (Pi + α
) is calculated. The magnitude of the average power Pi and the threshold value Ti are compared, and the determination result X is which one is larger. is output. Further, the calculated value (Pi+α) and the threshold value Ti are compared in magnitude, and a determination result X1 is output as to which one is larger. Input the judgment result X1, (Pi+α)
is greater than or equal to Ti, the threshold value Ti-H of the next interval is set equal to Ti, and (Pi+α)
is smaller than Ti, the threshold value Ti-H of the next interval
is set to (Pi+α). At the start of the operation, the threshold value Ti is a predetermined constant value T. The threshold value Ti is changed by the above-described method every time a unit of time elapses, and a determination result is output.

In this manner, the present invention is capable of determining the presence or absence of speech by automatically determining a threshold value according to the noise level while repeating the operations sequentially, without measuring the noise at a specific time.

"Embodiment" An input signal x (t) such as voice is converted by the AD converter 1 into a digital signal Xv sampled at a constant period (for example, every 125 μsec). This signal is subjected to sum-of-products calculation by a square calculation circuit 2, an addition circuit 3, and a delay circuit 4 (delay of 125 microseconds). Since the timing generator 5 zeros out the contents of the delay circuit 4 at fixed time intervals (one frame, for example 32 msec), an output proportional to the total power of the input signal in that interval is output from the adder circuit 3. This output is converted via a logarithmic circuit 6 into a logarithmic average power Pi within the interval.

The AD conversion circuit 1, the addition circuit 3, and the timing generation circuit 5 can each be realized using known techniques. For example, the square calculation circuit 2 and the logarithm circuit 6 are read-only memos! This can be realized by storing output values corresponding to inputs in an ICROM or the like. In this way, the logarithmic average power Pi within the interval is obtained at regular intervals, and is input to the section that determines the presence or absence of audio below.

A constant value α is added to the logarithmic average power Pi by an adding circuit 7,
It is compared with a threshold value Ti in a comparator circuit 8, and the determination result Xl controls a selection circuit 9. At this time, the comparison circuit 8 outputs l when Pi+α is greater than or equal to Ti, and outputs 0 when Pi+α is smaller than Ti. The selection circuit 9 selects an input terminal according to the logic shown below in response to the input from the control input terminal S1. and Dl are output to the output terminal Q.

If 51=O then Q” Dt If 51=1, then Q”D.

This output is input to the delay circuit 10 (delay of 32 msec) and is used as the threshold value Ti-H for the next section. There is a delay when this detection circuit starts operating.

A predetermined constant value is set in the path 10, and thereafter the contents are updated at regular intervals (every 32 msec). On the other hand, in the comparator circuit 11, the logarithmic average power Pi and the threshold value Ti are compared, and when Pi is greater than or equal to Ti, it is 1 (indicating that there is a sound), and when Pi is smaller than Ti, it is 1 (indicating that there is a sound). 0 (represents silence)
Output.

FIG. 2 is a diagram explaining the operating principle of this voice detection circuit. This voice detection circuit calculates the average power Pi of the input signal within each interval (32 ms), and sets this to a threshold value T
When compared with i, when Pi≧Ti, and when sound Pi<Ti, it is determined that there is no sound. When Pi is small, the threshold value Ti-1-t is used for the judgment process of the next section according to the logic shown below.
Update the value of

When Pi≧T1-α, Ti+t=Ti (no change) When PlgTi-α, TiH=Pi+α (updated) However, at the start of operation, Ti is set to a predetermined value To.

′? FIG. 53 schematically shows how the threshold value Ti changes over time when this voice detection circuit is operated. It can be seen that the value of Ti changes with time and finally settles to a value slightly larger than the noise level. In order for this voice detection circuit to operate stably, it is necessary that the noise level after the start of operation is approximately constant and that the fluctuation width is not extremely large. The value of α is set slightly larger than the fluctuation width of the noise level, and the initial value of the threshold value T
. If you set it higher than the expected noise level.

Voice detection can be performed stably.

This speech detection circuit operates stably even when background noise levels vary widely. However, in the following cases, the error in the determination result may become large.

(1) When the noise is extremely small; (2) When the noise is sudden; and (3) When the noise level increases over time. In the following, means for dealing with these problems will be described.

In the case of (1), the fluctuation width of the noise level on the logarithmic axis C2 becomes relatively large. Therefore, if this exceeds a predetermined value α, the number of errors in determining the noise portion as sound increases rapidly. In this case, since the noise level is sufficiently low,
Voice detection is sufficient even if the value of Ti is not made very small. Therefore, such unstable operation can be prevented by setting a lower limit value for the value of Ti and controlling it so that it does not fall below that value.

In the case of (2), the error occurs because a section in which the average power exceeds the threshold value is considered to be audible, even if it is sudden. In order to prevent this, it is practical to temporally smooth the output of this voice detection circuit. Regarding smoothing processing, there are already published algorithms that make the final judgment by giving various constraints such as the minimum duration allowed for a sound section and the minimum duration allowed for a silent section, so these should be used. be able to. (For example, the above-mentioned Yoshikawa, Ueda: Theory of Shinshu, 1982-
In the case of Ring 319 (B-1051pp(3)), the threshold value Ti changes while searching for the minimum value of the average power, so if the noise level becomes smaller over time, it can be followed, but as time passes This is because it has the disadvantage of not being able to keep up with it when it grows.To prevent this, 'l
'1 (7) If the value is not updated, instead of making the threshold value Ti-4-t of the next interval equal to Ti, Ti-
This problem can be solved by setting H=Ti+ε and setting ε to an appropriate value (2). When long hitting sections are continuous, the value of Ti increases little by little over time, making it possible to follow the gradual increase in noise.

FIG. 4 shows an example of a configuration having the lower limit value of Ti described in (1) and the function of increasing Ti described in (3). The added parts are a comparator circuit 12 and an adder circuit 13, and the selection circuit 14 is controlled by both outputs of the comparator circuits 8 and 12, and has an input terminal. + Dl h Outputs any input of D2. Input terminal. A value obtained by adding Ti and ε by an adder circuit 13 is input to the input terminals D1. D2 has (Pi+α
) * Comparison circuit 12 to which Tmin is input respectively
Then, (Pi+α) and Tmin are compared, and (Pi+α
)≧Tmin, output S. is 1, (Pi+α)<Tmin
So So becomes O. The selection circuit operates according to the following logic.

If So= s1= o then Q = D2SO=
1, if St = 0, then Q = DISo = 81 = 1
Then Q=D.

Each of the functions described in the above embodiments can be realized as a program running on an electronic computer or microprocessor, or can be implemented as a program running on a digital signal processor CD5P.
) can also be realized.

"Effects of the Invention" As explained above, by using the speech detection circuit of the present invention, it is possible to detect speech contained in an input signal under conditions of different noise levels. Unlike conventional devices, it does not measure noise at a specific time, but rather updates the threshold according to the noise level while determining whether there is a sound or not. This saves time and eliminates the problem of errors in subsequent determinations due to voices, sudden noises, etc. occurring at the time of noise measurement.

In particular, in devices that detect voice via telephone lines, the noise level varies considerably depending on the line distance, connection route, etc. If the voice detection circuit of this invention is used, even if a person is speaking at the time the voice/silence determination operation is started, or even if a sudden noise occurs, the subsequent determination results will be extremely incorrect. The problem will go away.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing a first embodiment of the invention, FIG. 2 is a diagram showing the operating principle of the circuit of the invention, and FIG. 3 is a diagram schematically explaining the operation of the circuit of the invention. FIG. 4 is a block diagram showing a second embodiment of the invention. Patent applicant Nippon Telegraph and Telephone Corporation

Claims (3)

[Claims] (1) Divide the continuously input audio signal into units of time (hereinafter, the description will focus on the i-th divided section), and calculate the average power of that section (hereinafter expressed as P_i). means for calculating a value (P_i+α) obtained by adding a constant value α to the average power P_i; and determining which one is larger by comparing the magnitudes of the average power P_i and a threshold value (hereinafter referred to as T_i). Result X_0
means for outputting the calculated value (P_i+α) and the threshold value T_i and outputting a judgment result X_1 to determine which is larger;
If it is greater than or equal to, set the next interval threshold (T_i+1) equal to T_i and (P_i+α
) is smaller than T_i, the next interval threshold T_
means for setting i+1 to (P_i+α), and at the time of starting the operation, the threshold value T_i is set equal to a predetermined constant value T_0, and each of the above-mentioned means is set every unit time. A voice detection circuit characterized in that it outputs a determination result X_0 by changing a threshold value T_i. (2) Input the judgment result (X_1) and (P_i+α
) is greater than or equal to T_i, set the threshold value (T_i+1) of the next interval equal to T_i, and (
P_i+α) is smaller than T_i, and (P_i+α)
is larger than the constant value T_m_i_n, the threshold value (T_i+1) of the next section is set to (P_i+α), and (
P_i+α) is smaller than T_i, and (P_i+α)
2. The voice detection circuit according to claim 1, further comprising means for setting a threshold value (T_i+1) for the next section to T_m_i_n when T_m_i_n is smaller than a constant value T_m_i_n. (3) Input the judgment result (X_1) and (P_i+α
) is larger than or equal to T_i, the threshold value (T_i+1) of the next interval is set to a value larger than T_i by a constant value ε, characterized in that the method is characterized in that voice detection circuit.