JPH0713585A - Speech section segmentation device - Google Patents

Abstract

PURPOSE:To provide the speech section segmentation device which correctly segments a speech section from an input signal containing a speech. CONSTITUTION:The input signal containing the speech is stored in a data buffer 13 and an amplitude threshold value At is found on the basis of the noise amplitude in a voiceless section and stored in an amplitude threshold buffer 17; and a count threshold value Nt is found on the basis of how many times the input signal in the voiceless section exceeds the amplitude threshold value At and stored in a count value threshold buffer 21. An amplitude comparison part 16 compares the input signal stored in the data buffer 13 with the amplitude threshold value At and a counter 19 counts how many times the input signal exceeds the amplitude threshold value At; and a count value comparison part 20 compares the count value N with the count threshold value Nt and the speech section of the input signal is determined according to the comparison result.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice section cutting device for cutting a voice section for voice recognition.

[0002]

2. Description of the Related Art Conventionally, a human being has been made to perform a predetermined operation by directly operating a handle or the like, but recently, by directly inputting a voice to the machine. Attempts have been made to give a command to perform a predetermined operation, and various voice recognition devices have been developed for this purpose.

Conventionally, as a voice recognition device of this type, when a voice input is input from a microphone 1 as shown in FIG. 9, this voice input is A / D converted by an A / D converter 2.
After the conversion, the converted data is given to the voice section cutout unit 3, the voice section is cut out, and the acoustic analysis unit 4 continuously performs acoustic analysis such as frequency analysis on these sections and parameterizes them and gives them to the comparison unit 5. There is a method in which the most similar one is output as a recognition result to the display unit 7 by comparing with the dictionary parameter of the parameter dictionary 6 prepared in advance.

The recognition result in this case varies depending on the voice segment to be cut out. That is, if the voice segment to be cut out is different, the value of the parameter is naturally different, and if the correct voice segment cannot be cut out, erroneous recognition will occur.

However, when the voice section is actually cut out from the voice data, it is difficult to correctly cut out the voice section because the voice input from the microphone 1 includes noise around the speaker. Sometimes.

Therefore, conventionally, as a voice segment cutout device, as shown in FIG.
A / D converter 32 using the clock signal from
Convert and write this data in the buffer memory 33.
Further, the output of the frame data counter 34 outputs an address of a predetermined time interval T (one frame) unit from the address generation unit 35, reads the audio data from the buffer memory 33, and supplies the audio data to the amplitude comparison unit 36. Then, the amplitude comparison unit 36 compares the amplitude threshold value calculated by the amplitude threshold value calculation unit 37 from the output (ambient noise) of the buffer memory 33 in the non-voice section in advance and stored in the threshold value buffer 38, and the comparison is made. There is one that detects the voice section from the result.

Then, in such a voice segment cutting device, as shown in FIG. 11, when the input signal I is given to the amplitude threshold S which is set in advance based on the noise in the non-voice segment, the input signal I The point at which the amplitude first exceeds the threshold value S is the beginning point a of the voice, and the point at which the amplitude of the input signal I finally exceeds the threshold value S is b. The end point c is set, and the section between the points a and c is set as the voice section.

[0008]

However, in such a voice segment cutout device, since there is a large variation in the relation that the amplitude threshold S is set based on the ambient noise in the non-voice segment, it is assumed that FIG. ), If the threshold value S is set to a large value and a voice signal having a small amplitude at the first portion is given, this portion may not exceed the threshold value S, and the starting point a of the original voice section is The point a ′ after this may be erroneously detected as the start point of the voice, and the correct voice section may not be cut out. If the threshold value S is set to be small as shown in FIG. Since the noise part which is not the voice signal exceeds the threshold value S, the starting point a of the original voice section is erroneously detected as the starting point a of the voice, and the correct voice section cannot be cut out in this case as well. Question There was a point.

The present invention has been made in view of the above circumstances, and an object of the present invention is to provide a voice section cutout device which can correctly cut out a voice section from an input signal containing a voice.

[0010]

According to the present invention, a storage means for storing an input signal containing a voice, a comparing means for comparing the input signal with an amplitude threshold value, and a number of times the input signal exceeds the amplitude threshold value are counted. Counter means, a count value comparison means for comparing the count value of the counter means with a count threshold value, and a voice section determination means for determining the voice section of the input signal from the comparison result of the count value comparison means. There is.

Further, the threshold value is characterized in that the amplitude threshold value is set by adding a predetermined value to the average amplitude of noise in the non-voice section of the input signal. Further, the count threshold is obtained by dividing the non-voice section into predetermined time sections, counting the number of times the input signal exceeds the amplitude threshold for these time sections, and adding a predetermined value to the average of these count values. Is set.

Further, the voice section determining means is characterized in that the voice section of the input signal is determined on condition that the same state is continuously obtained a plurality of times as the comparison result of the count value comparing means.

[0013]

As a result, according to the present invention, the amplitude threshold is set based on the noise amplitude of the input signal including voice in the non-voice section, and the number of times the input signal in the non-voice section exceeds the amplitude threshold. By using the count threshold set by the above, the voice section of the input signal is determined from the number of times the input signal exceeds the amplitude threshold and the comparison result of the count threshold. As a result, the number of times that the input signal exceeds the amplitude threshold in a portion other than the voice section can be clearly reduced as compared with the voice section, and it is possible to prevent a noise part other than the voice section from being erroneously detected as the voice section.

[0014]

Embodiments of the present invention will be described below with reference to the drawings. (First Embodiment) FIG. 1 shows a schematic structure of the first embodiment. In the figure, reference numeral 11 is an A / D converter, which is supplied with an input signal including voice, and this input signal is converted into an A / D signal by using a clock signal of a clock generator 12.
The data is D-converted and this data is written in the buffer memory 13.

The clock signal of the clock generator 12 is also given to the frame data counter 14. The frame data counter 14 has a predetermined time interval T (1
The frame data counter 14 counts the number of frames, and the output of the frame data counter 14 causes the address generator 1
An address is generated in 5, and the data is read from the buffer memory 13. In this case, in the buffer memory 13, in the setting of the threshold value described later, the data is read by specifying the address at intervals of the time section T, and when determining the start point or the end point of the voice section, the time section T is set as the time axis. The data is read out while the address is specified while shifting for a predetermined time along.

It should be noted that the frame data counter 14 clears its contents every time it counts the data in the time section T (1 frame). Buffer memory 1
The data read from No. 3 is given to the amplitude comparison unit 16. The amplitude comparison unit 16 compares the data A read from the buffer memory 13 with the amplitude threshold At stored in the amplitude threshold buffer 17, and when there is data A exceeding the threshold At (A> At). To generate output.

Here, the amplitude threshold At of the amplitude threshold buffer 17 is calculated by the amplitude threshold calculator 18 based on the data (ambient noise) from the buffer memory 13 in the non-voice section.

The output of the amplitude comparison section 16 is given to the counter 19. The counter 19 counts the number of times the output (A> At) from the amplitude comparing section 16 is generated. Further, the counter 19 is cleared by the output of the frame data counter 14 for each time period T.

The count value of the counter 19 is given to the count value comparing section 20. This count value comparison unit 20
Is for comparing the count value N of the counter 19 with the count threshold value Nt stored in the count value threshold value buffer 21. When the comparison result is N> Nt, the count value Nt is output as a voice section.

In this case, the count value threshold value Nt of the count value threshold value buffer 21 is obtained by the count value calculating section 22 based on the count value (ambient noise) of the counter 19 in the non-voice section.

Next, the operation of the embodiment configured as described above will be described. First, the amplitude threshold value At and the count threshold value Nt are set according to the flowchart shown in FIG.

In this case, in the non-voice section containing no voice, the output from the buffer memory 13 is set to the amplitude threshold calculation unit 1
8 to obtain the average amplitude An of the noise (step 20
1) Then, + α is added to this average amplitude An to determine the amplitude threshold At (step 202), and the amplitude threshold At is written in the amplitude threshold buffer 17. Further, the non-voice section of the buffer memory 13 for which the above-mentioned average amplitude An is obtained is divided into a plurality of sections by the time section T, and the amplitude comparison unit 16 compares the data A of each time section T with the amplitude threshold At to obtain the data A. Exceeds the threshold At>A>
The number of times At is counted by the counter 19 to obtain the count value N, and the count value N is given to the count value calculation unit 22 to obtain the average count value Nn for each time interval T (step 203). + Β is added to the average count value Nn to determine the count threshold value Nt (step 204), and the count value threshold value buffer 21 is written.

Next, when a voice section including a voice is given, the start point of this voice section is determined by the flowchart shown in FIG. In this case, the data A relating to the first time section T is read from the buffer memory 13 and compared with the amplitude threshold At in the amplitude comparison unit 16. Then, the number N of times the data A exceeds the amplitude threshold At is counted by the counter 19 (step 301), and this counter 1
The count value of 9 is given to the count value comparison unit 20 and compared with the count threshold value Nt (step 302). If the comparison result is N> Nt, the start point of the time section T is determined as the start point of the voice section (step 30).
3). On the other hand, when the comparison result is N ≦ Nt, the corresponding time section T is shifted by a predetermined time along the time axis (step 304), the process returns to step 301, and the same operation as described above is repeated. After that, a time section T in which N> Nt is obtained appears, waits for the start point of the voice section to be determined, and then the processing ends.

Next, the determination of the end point of the voice section is executed according to the flowchart shown in FIG. In this case, assuming that the determination of the start point of the voice section is confirmed (step 401), the counter 19 counts the number N of data A exceeding the threshold value At for the time section T immediately after this (step 402). . And this counter 1
The count value of 9 is given to the count value comparison unit 20 and compared with the count threshold value Nt (step 403). If the comparison result is N≤Nt, the end point of the time section T is determined as the end point of the voice section (step 40).
4). On the other hand, when the comparison result is N> Nt, the corresponding time section T is shifted by a predetermined time along the time axis (step 405), the process returns to step 402, and the same operation as described above is repeated. After that, a time section T in which N≤Nt is obtained appears, waits for the end point of the voice section to be determined, and then the processing ends.

Next, a description will be given of how such a voice segment cutout device is used to set each threshold value and determine the start point of the voice segment for an actual input signal. Figure 5
Indicates an input signal of a voiceless section which does not include voice. In this case, the average amplitude An of the data (noise) A is obtained from the input signal of the voiceless section, and the average amplitude An is + α.
Is added to determine the amplitude threshold At. Further, the non-voice section is divided into M sections N0 to NM-1 in the time section T, and the number N of times when the data (noise) A exceeds the amplitude threshold At is counted for these sections N0 to NM-1. Is averaged over M time sections T to obtain an average count value Nn per one time section T, and this average count value Nn is + β.
Is added to determine the count threshold Nt.

Next, FIG. 6 shows an input signal of a voice section including a voice. In this case, in the first time section T0, the number N of data A exceeding the amplitude threshold At is counted, and this count value is counted. Compare N with count threshold Nt.
If the comparison result is N ≦ Nt, the time section T0 is shifted by a predetermined time along the time axis to newly set the time section T1, and the amplitude threshold value A is set for the time section T1.
The number N of data A exceeding t is counted, and this count value N is compared with the count threshold Nt. If the comparison result here is N≤Nt, the time interval T1 is further shifted to newly set the time interval T2, and the above-described operation is repeated. If N> Nt is obtained in the time section Ti, the start point of the time section Ti is determined as the start point of the voice section.

Accordingly, in this way, the amplitude threshold At set based on the noise amplitude of the input signal including voice in the non-voice section and the amplitude threshold At of the input signal in the non-voice section are set.
On the basis of the count threshold value Nt set based on the number of times over, the voice section of the input signal is determined from the comparison result of the number N of times the input signal exceeds the amplitude threshold value At and the count threshold value Nt. Since the number of times the input signal exceeds the amplitude threshold can be made significantly smaller than that in the voice section, the noise part other than the voice section is erroneously detected as the voice section as compared with the conventional method in which only the amplitude threshold is set. This can be reliably prevented, and the voice section can be correctly cut out from the input sound signal.

(Second Embodiment) FIG. 7 shows that if a relatively large noise NS exists before the start of a voice section, this noise N
It shows a case where S may be erroneously recognized as the beginning of a voice section.

Therefore, in order to eliminate such inconvenience, in the second embodiment, the count value N of the data A exceeding the amplitude threshold At is N> Nt with respect to the count threshold Nt in the first time section T0. However, without immediately determining the starting point, the time interval T0 is shifted to newly set the time interval T1, and such an operation is repeated, and N> Nt.
The starting point of the voice section is determined only when the relationship of is obtained a predetermined number of times in succession. In the case of FIG. 7, the relation of N> Nt is continuous three times in the noise NS portion, but if the number of consecutive times is set to four or more in advance, it is possible to prevent erroneous recognition as the start of the voice section. it can.

(Third Embodiment) FIG. 8 shows a case where a signal portion having a small amplitude exists in the middle of a voice section, which may be erroneously recognized as the end of the voice section.

Therefore, in order to eliminate such inconvenience, in the third embodiment, the count value N of the data A exceeding the amplitude threshold At is N≤Nt with respect to the count threshold Nt in the first time section T0. Even if the end point is not determined immediately, the time interval T0 is shifted and a new time interval T is added.
By setting 1 and repeating such an operation, the end point of the voice section is determined only when the relationship of N≤Nt is continuously obtained a predetermined number of times. In the case of FIG. 8, the relationship of N ≦ Nt is continuous twice in the signal portion with a small amplitude, but if the number of continuous times is set to three times or more in advance, it may be erroneously recognized as the end point of the voice section. Can be prevented.

The present invention is not limited to the above-mentioned embodiments, and can be carried out by appropriately modifying it without changing the gist.
For example, in the above-described embodiment, the threshold value for noise is set by the absolute value of the amplitude of the input signal. However, the threshold value for noise is set by using the logarithmic value of the input signal and converting it into power. You may

[0033]

According to the present invention, the amplitude threshold value is set based on the noise amplitude of the voice-free input signal in the non-voice section and the number of times the input signal in the non-voice section exceeds the amplitude threshold value. Since the count threshold is set to determine the voice section of the input signal from the number of times the input signal exceeds the amplitude threshold and the comparison result of the count threshold, the number of times the input signal other than the voice section exceeds the amplitude threshold is Since it can be clearly reduced compared to the voice section, it is possible to prevent a noise part other than the voice section from being erroneously detected as the voice section, and the voice section can be correctly cut out from the input sound signal. Further, by doing so, the threshold value can be set small even for a voice that starts from a consonant having a relatively small amplitude, so that a correct voice segment can be cut out.

[Brief description of drawings]

FIG. 1 is a diagram showing a schematic configuration of an embodiment of the present invention.

FIG. 2 is a flowchart for explaining the operation of the embodiment.

FIG. 3 is a flowchart for explaining the operation of the embodiment.

FIG. 4 is a flowchart for explaining the operation of the embodiment.

FIG. 5 is a diagram for explaining how threshold values are set for an actual input signal.

FIG. 6 is a diagram illustrating how a start point of a voice section is determined with respect to an actual input signal.

FIG. 7 is a diagram for explaining the second embodiment of the present invention.

FIG. 8 is a diagram for explaining a third embodiment of the present invention.

FIG. 9 is a diagram showing a schematic configuration of a conventional voice recognition device.

FIG. 10 is a diagram showing a schematic configuration of a conventional voice segment cutout device.

FIG. 11 is a diagram for explaining a conventional voice segment cutout device.

FIG. 12 is a diagram for explaining a conventional voice segment cutout device.

[Explanation of symbols]

11 ... A / D converter, 12 ... Clock generator, 13 ... Buffer memory, 14 ... Frame data counter, 15 ...
Address generation unit, 16 ... Amplitude comparison unit, 17 ... Threshold buffer, 18 ... Amplitude threshold calculation unit, 19 ... Counter, 20 ... Count value comparison unit, 21 ... Count value threshold buffer, 22
… Count value calculator.

Claims (4)

[Claims] 1. A storage unit for storing an input signal including voice, a comparison unit for comparing the input signal with an amplitude threshold value, a counter unit for counting the number of times the input signal exceeds the amplitude threshold value, and a counter unit. 5. A voice segment cutout device, comprising: a count value comparing means for comparing a count value with a count threshold value; and a voice segment determining means for determining a voice segment of the input signal from a comparison result of the count value comparing means. 2. The speech segment extraction device according to claim 1, wherein the threshold is set by adding a predetermined value to an average amplitude of noise in a non-speech segment of an input signal. 3. The count threshold value divides a non-voice section into predetermined time sections, counts the number of times the input signal exceeds the amplitude threshold value in these time sections, and adds a predetermined value to the average of these count values. The voice segment extraction device according to claim 1, wherein the count threshold value is set. 4. The voice section determining means determines the voice section of the input signal on condition that the same state is obtained a plurality of times consecutively as a comparison result of the count value comparing means. Voice segment cutting device.