JPH0247698A - Speech section detection system - Google Patents

Abstract

PURPOSE:To reduce errors in speech section detection even in the presence of a loud noise by finding the mean noise level value of all acoustic power values of acoustic power sampled at the time of background noise level measurement except a specific quantity from the maximum value and a specific quantity from the minimum value and mean noise dispersion. CONSTITUTION:A microprocessor 30 as a threshold calculation part calculates the mean noise level NL' of all acoustic power values P1 except the specific number Nmax of values from the maximum value and the 2nd specific number Nmin of values from the minimum value in order and the mean noise dispersion ND', and then calculates a speech segmentation level VL from the both. Then the level is sent to a speech section detection part 16 through a system bus 36 under the control of the microprocessor 30. Consequently, the errors in speech section detection can be reduced in the noise environment of a click noise, etc.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a method for detecting speech intervals in a speech recognition device.

(Prior Art) A typical speech recognition device performs a process of detecting a section in which speech exists from an input acoustic signal (hereinafter referred to as speech section detection processing), and a process of recognizing and determining the content of the detected speech (hereinafter referred to as speech section detection processing). It can be roughly divided into two types of processing, called recognition processing and !fi) processing.

Generally, in order to perform such an operation, a speech recognition device analyzes an input acoustic signal for each minute time period called an acoustic frame and calculates its characteristic parameters. Typical characteristic parameters include acoustic power and power spectrum.

Speech section detection utilizes the property that a speech section has higher acoustic power than other sections.

An example of such a conventional voice section detection processing method is disclosed in Japanese Patent Application Laid-Open No. 60-114900. An example of the configuration of this conventional method will be explained with reference to FIG.

A zero-order power calculation unit 14 that samples an acoustic signal input from an external input unit 10, such as a microphone or a telephone, in the A/D conversion unit 12 and converts it into a digital signal sequence.
Then, the acoustic power PI (I indicates a frame number) is calculated for each frame from this digital signal sequence (hereinafter simply referred to as an input signal), and is sent to the voice section detection section 16 and the threshold value setting section 18, respectively. In the threshold 1 setting section 18, as will be described later, the average noise level is calculated based on this acoustic power P, and sent to the speech section detection section 16, and in this speech section detection section 1δ, the acoustic power P and the average noise level are calculated. In the zero-order recognition unit 20, which detects and determines the voice section from the level, recognition processing is performed on the voice bang consisting of the acoustic power series of the voice section, and the recognition result is sent to the external device 22, for example, a computer. and other necessary display devices.

In the conventional voice recognition snowmaking system having such a configuration, prior to the recognition operation, the background noise level is measured for the purpose of setting the average noise level for voice section detection as described above. This is to measure the properties of acoustic power in a no-input state and determine an appropriate threshold for voice section detection.

This process will be explained below. Based on the acoustic power P obtained by the power calculation section 14 from the acoustic signal inputted from the external input section 10, the threshold setting section calculates the average noise level NL and the average noise variance N for one day. Calculate. These average noise level N and average noise fraction ND are given by the following equations (1) and (2), respectively, where N is the number of measurement frames.

Roughly speaking, the voice extraction level VL has been determined from the average noise level NL and the average noise fraction ND according to the following equation (3).

VL = NL + NI
・ ・ ・ ・ ・ ・ ・ ・ (3) Here, N1
is a count determined in advance by the system, and usually takes a value of about 2 to 4. The voice cutting level VL calculated in this way is used by the voice section detection unit 16 thereafter.

Next, a conventional voice section detection operation will be briefly explained.

First, as usual, an acoustic signal input from the external input section 10 is converted into an input signal by the A/D conversion section 12, and then the acoustic power P is calculated by the power calculation section 14. An example of this acoustic power P is shown in FIG.
In the diagram, the vertical axis represents the acoustic power P and the horizontal axis represents the frame number. In the diagram, the broken line represents the audio cutting level VLt. , and , are the voice start and voice end of the voice section. Further, V, ,V are a voice start frame and a voice end frame, and the frame period is usually about 8 milliseconds.

The voice section detecting section 16 performs the process of cutting out the voice section described above, and conventionally, the first frame in which the following conditions (1) to (2) are satisfied for the sound power P1 is set as the starting frame of the voice section.

■When a frame I with which the starting edge condition P≧V exists continues for a number of frames, that is, N2 frames determined in advance by experience, this frame E is changed to the starting edge frame V.
, and so on.

(2) Termination condition Also, after detecting the start frame V, the frame immediately before the frame in which the following conditions are satisfied for the first time is defined as the end frame V of the voice section.

When frames such that P<V continue for a plurality of frames, that is, N3 frames determined in advance by experience, after frame E.

■Exclusion conditions In addition, if the voice section length V L E H meets the following conditions, it will not be considered a voice section.

VLEII<N, or VLEN>Ns, where VLE11=Vε-S+1, and N4 and N5 are the numbers of frames predetermined by experience.

(Problems to be Solved by the Invention) The conventional calculation of the audio cutout level VL described above assumes that the distribution of acoustic power of background noise is close to a normal distribution. In fact, this approximation holds true in a quiet environment. However, if you are in an environment where the noise level is high, or if you are using a telephone, etc. ! In the case of input conditions that can be obtained via , there is noise such as a click sound that has a short duration but has an extremely high peak acoustic power, so this approximation often deviates from this approximation. Similarly, when the sound power level is considerably high, the distribution increases.

Therefore, if such noise occurs when measuring the background noise level, the average noise level NL and the average noise variance N
Both D are calculated to be high, which causes voice section detection errors. One method to reduce this decrease is to lengthen the average noise level measurement time N, but this method increases the preparation time before starting recognition and reduces the responsiveness of the voice recognition snowmaking itself. Therefore, a sufficient measurement time N could not be adopted.

An object of the present invention is to provide a voice section detection method that can set a voice segmentation level VL that can significantly reduce voice section detection errors even in a noisy environment such as the above-mentioned click sound.

(Means for Solving the Problems) In order to achieve this object, according to the speech section detection method of the present invention, the threshold value calculation section selects the first sound power P1 in order from the one having the largest value. The average noise is calculated for all the remaining acoustic powers P, excluding the acoustic powers of a predetermined number N max and the acoustic powers of a second predetermined number N m I n in order of the smallest value. After calculating the level NL' and the average noise variance No', the average noise level NL' and the average noise variance N are calculated. The method is characterized in that the audio cutout level L is calculated from .

(Function) With this configuration, the original sound power distribution when no voice is input, excluding the high sound power side caused by noise such as click sound and the low sound power side caused by other noises, can be obtained. Since this method determines the audio extraction level VL using the acoustic power in the intermediate distribution region where the acoustic power is concentrated, it is possible to determine the appropriate audio extraction level VL8 extremely easily without being affected by noise components with high peak power. . As a result, errors in voice section detection are reduced. Therefore, a voice recognition bag M with excellent overall recognition performance is provided.

(Example) Hereinafter, an example of the voice section detection method of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram for explaining an embodiment of the voice section detection method of the present invention, and FIG. 5 is a flowchart of processing in a dark value setting section.

In FIG. 1, the same constituent components as those shown in FIG. 2 are denoted by the same reference numerals (), and detailed explanation thereof will be omitted.

Further, in FIG. 1, 24 is a threshold value setting section corresponding to the conventional threshold value setting section 18 shown in FIG. 2, but it differs from this conventional threshold value setting section 18 in function and internal configuration.

First, the threshold value setting section 24 in this embodiment will be explained with reference to FIG. 5.

In this embodiment, first, each frame I
The power calculation unit 14 calculates the acoustic power P (I) for each (I=1, . . . , N), and sends this to the dark value setting unit 24 and the voice section detection unit 16.

In the threshold value setting section 24, under the control of the microprocessor 30, these acoustic powers P (I) are transferred from the power calculation section 14 via the system bus 36 to each memory area 1 of the memory 32, MEM (1), RMEM ( 2), R.M.
・・Memorize - hours in PMEM (N). In this case, I=
1. Start processing from the (first) frame (step 51). 6. Next, determine whether it is an INN. Step S2
), and when I≦N, the acoustic power P of the first frame is stored in the memory area vtRMEM (1) (step 53) 0 Next frame number I! Proceed to the next I=2 (step S4), return to the above-mentioned step S2,
Second (I=2) after performing steps S2 and S3
Sound power of the frame P2t Memory area tSiRMEM
(2) - Time memorization. In this way, sequentially, I=N
Up to each acoustic power P, the corresponding memory area R
To MEM (N) - Time memorization.

In step S2, if it is determined to be INN, the acoustic powers P1 to PN stored in each memory area RMEM(1) to RMEM(N) of the memory 32 are sorted in ascending order under the control of the microprocessor. The result is sent to the work memory 34 via the system bus 36, and the memory area SMEM (1
), SMEM (2) 5. ,, SMEM (N) in order of size (step S5). Therefore, for example, the memory area SMEM(1) stores the acoustic power P1 having the smallest peak value, and conversely, the memory area SMEM(N) stores the acoustic power P1 having the largest peak value. That is, in this embodiment, the memory area S
The following relationship holds true for the magnitude of the acoustic power Pl stored in MEM(J) (J=1, . . . , N).

SMEM(+) ≦SMEM(2) ≦...SM
EM(N) . . . (4) and all the remaining acoustic powers P1 except for the acoustic powers P corresponding to these numbers are read out from the work memory 34 to the microprocessor 3o (step S6).

Next, in the microprocessor 30, the following equation (5) (
Accordingly, the average noise level N,' is calculated and the result is temporarily stored in the memory of the microprocessor 3o (step 57).6 Next, the microprocessor 30
Then, calculate the average noise level N,' expressed by the following equation.

For this purpose, a memory (not shown) of the microprocessor 30 stores a predetermined number N of sound levels counted from the maximum sound power in ascending order and not used for calculating this average noise level, which is predetermined by experience. □ and the number of sound levels N Mll'l that are not used in calculating this average noise level, counted in order from the minimum sound power to the highest one, which are also predetermined by experience, are stored, and these are stored. N 1lla The average noise variance N. ' is calculated, and the result ND' is temporarily stored in the memory (step S8).

Next, these average noise levels N,' and average noise variances N,
D' and the coefficient N+, which has been previously determined based on experience and stored in the memory of the microprocessor 30, are read out and the audio cutting level VL is determined according to the following equation (7).
' is determined (step S9).

NmIn should be set to an appropriate value depending on the probability of occurrence of peak noise and the nature of duration. 0 Usually N□8 is 1/10 to 11 of N of the number of measurement frames.
50, and N m L n is preferably set to a value of 1/1° to 1150 to 0 of N.

The voice section detection process and the recognition process are the same as in the conventional example, so their explanation will be omitted.

It is clear that the embodiments described above are only preferred examples of the invention, and that the invention is not limited only to the embodiments described above.

VL = NL+ NIX No. ・ ・ ・ ・
・ ・ ・ ・ ・ ・ ・ (7) When the threshold value setting unit 24 completes the processing of steps S] to S9 described above, the resulting audio cutout level VL′ is set as the audio cutout level VL′ via the system bus 36 under the control of the microprocessor 30. It is sent to the section detection section 16. The measurement time N is usually 0.16 to 32 seconds, and when the frame period is 8 milliseconds, N = 20 to 4 seconds (effects of the invention).As is clear from the above explanation. According to the voice section detection method of the present invention, N maX acoustic powers are selected from the largest value among the acoustic powers P that have been suppressed when measuring the background noise level, and N maX acoustic powers are suppressed from the one that has the smallest value. In order, N, , 1. By determining the average noise level value NL' and average noise variance ND of all the remaining acoustic powers P except for the acoustic powers of ,
Since it is configured so that an appropriate speech extraction level can be set, speech section detection errors are extremely reduced even under high noise, and thus the recognition section M can be realized with excellent overall recognition performance. .

20... recognition section, 24...Threshold value setting section, 32...Memory, 36...System bus.

22...External device 30...Microprocessor 34...Work memory

[Brief explanation of the drawing]

Fig. 1 is a block diagram for explaining the speech interval detection method of the present invention, Fig. 2 is a block diagram for explaining the conventional speech interval detection method, and Fig. 3 is a block diagram for explaining the speech interval detection method of the present invention and the conventional method. FIG. 4 is a diagram showing an example of the sound power distribution, and FIG. 5 is a flowchart of the operation of the calculation process of the audio cutout level.

Claims (1)

[Claims] (1) The power calculation unit calculates the acoustic power P for each minute time period called a frame from the input acoustic signal from the external input unit.
_1 is calculated, and the acoustic power P_1 is calculated in the threshold value setting section.
Calculate the average noise level based on the acoustic power P
A speech recognition device configured to detect a speech section from _1 and an average noise level, perform recognition processing on a speech pattern determined by the speech section in the recognition section, and output the result to an external device. In detecting the section, the power calculation section measures the acoustic power P_1 for a predetermined period of time with no audio input, and the threshold calculation section measures the acoustic power P_1 having the largest value among the acoustic powers P_1. Average noise for all the remaining acoustic powers P_1 excluding the first predetermined number of acoustic powers N_m_a_x in order from the lowest value and the second predetermined number of acoustic powers N_m_i_n in the order of the smallest value. Level N_
A voice section detection method characterized in that after calculating the average noise level N_L' and the average noise variance N_D', a voice segmentation level V_L is calculated from the average noise level N_L' and the average noise variance N_D'.