JP5342629B2 - Male and female voice identification method, male and female voice identification device, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To accurately identify sexuality of a speaker of a speech signal even when time length of the input speech signal is extremely short. <P>SOLUTION: A method for identifying male/female voice includes: extracting a speech feature amount from an input speech signal; and identifying the sexuality of a speaker of the speech signal on the basis of likelihood obtained by collating the speech feature amount with a male speech acoustic model and a female speech acoustic model. When the time length of the speech signal is less than a predetermined time length L, the speech signal is repeatedly extended till becoming to have the time length L or more, the speech feature amount is extracted using the extended speech signal, and a collation and an identification are performed using a speech recognition grammar corresponding to the repeat. <P>COPYRIGHT: (C)2013,JPO&amp;INPIT

Description

  The present invention relates to a male / female voice identification method, a male / female voice identification device, and a program for identifying the gender of a speaker of an input voice signal.

  The male / female voice identification technique for identifying the gender of a speaker from the input voice signal is an important technique not only for identifying gender but also for improving the accuracy of voice recognition, for example.

  Up to now, when the input audio signal identifies male voice or female voice, the voice feature is extracted from the voice signal, and the extracted voice feature is statistical model such as GMM (Gaussian Mixture Model). The likelihood is obtained by collating with the acoustic model for male voice and the acoustic model for female voice created based on the conversion, and the male voice and female voice are identified based on the likelihood.

  Patent Document 1 describes that the speech feature amount extracted from the speech signal input in this manner is collated with an acoustic model for male voice and an acoustic model for female voice, and the likelihood is obtained. Male voices and female voices can be identified based on this.

  FIG. 11 shows an example of the configuration of a gender voice identification device for identifying the gender of a speaker of an input voice signal by the method as described above. The male / female voice identification device includes a male / female voice identification processing unit 10, a male voice acoustic model 20, and a female voice acoustic model 30. The male voice acoustic model 20 includes a voice section model 21 and a non-speech section model 22, and similarly, the female voice model 30 includes a voice section model 31 and a non-speech section model 32.

  In this example, the male / female voice identification processing unit 10 includes a voice feature amount extraction unit 11, a recognition grammar setting unit 12, and an identification unit 13. The voice feature quantity extraction unit 11 extracts the voice feature quantity of the input voice signal (digital voice signal that has been A / D converted and digitized). The recognition grammar setting unit 12 collates the voice feature quantity with the male voice acoustic model 20 and the female voice acoustic model 30 and sets the recognition grammar used when obtaining the likelihood. The identification unit 13 uses the recognition grammar, compares the voice feature amount with the male voice acoustic model 20 and the female voice acoustic model 30, obtains the likelihood, and identifies the gender of the speaker of the voice signal based on the obtained likelihood. The male / female voice identification processing unit 10 outputs the result of identification in this way.

  Since it is necessary to identify male voices and female voices in units of utterances, the recognition grammar set by the recognition grammar setting unit 12 is generally a recognition grammar (1) as shown below. Note that the recognition grammar (1) below is described in an expanded form of the BNF notation.

・ Recognition grammar (1)
$ [p] = pause;
$ [g] = garbage;
$ START = $ p $ g $ p;
Here, $ [xxx] = means a symbol declaration, pause on the right side is a symbol representing non-speech such as silence, and garbage is a symbol representing speech. $ START is a start symbol that represents the entire sentence. Note that the symbol = represents the definition, the symbol [] represents the word notation, and the symbol; represents the end of the definition.

  The recognition grammar (1) is assumed to appear in the order of non-speech → speech → non-speech.

JP 2011-13543 A

  By the way, for example, when male and female voices are identified for a very short audio signal of 1 second or less, the following problems occur.

  That is, when extracting a voice feature value from an input voice signal, generally, for example, CMN (Cepstrum Mean Normalization), CVN (Cepstrum Variance Normalization), or the like is used to remove the bias of the voice feature value by a speaker. Performs normalization processing of voice feature. However, since such normalization processing is based on processing based on statistical analysis of speech features, if a speech signal of a certain length is not input, a statistically correct result cannot be obtained, resulting in correct normalization. May not be able to be processed.

  Therefore, for example, when a very short speech such as “Yes” or “No” is input, the bias depending on the speaker remaining in the speech feature amount extracted and normalized from the speech signal is / There is a possibility that problems such as the influence of the female models on the acoustic model may be affected and the identification accuracy may be lowered.

  In view of such problems, the object of the present invention is to make it possible to accurately identify the gender of the speaker of the audio signal even when the time length of the input audio signal is very short. An object is to provide a discrimination method and a gender voice discrimination device.

  According to the first aspect of the present invention, the voice feature is extracted from the input voice signal, and the gender of the speaker of the voice signal is determined based on the likelihood that the voice feature is collated with the male voice model and the female voice model. In the male / female voice identification method for identifying, when the time length of the audio signal is less than the predetermined time length L, the audio signal is repeatedly expanded until the time length becomes equal to or longer than the time length L, and the audio feature amount is used using the expanded audio signal. Are extracted, and the collation and identification are performed using a recognition grammar corresponding to the repetition.

  According to a second aspect of the invention, in the first aspect of the invention, when the repetition is performed, a voice section of the voice signal is detected, and the voice signal is expanded by repeating only the voice section.

  In the invention of claim 3, in the invention of claim 2, when the length of the detected speech section is less than the threshold value T, the speech section is rejected and the identification is not performed.

  According to a fourth aspect of the present invention, in any one of the first to third aspects, the time length L can be set from the outside.

  According to a fifth aspect of the present invention, in any one of the first to third aspects, the time length L is calculated from a computer load for executing gender discrimination and a required response time.

  According to the invention of claim 6, the gender voice identification device determines whether or not the time length of the input audio signal is less than the predetermined time length L. When it is output to the decompression unit and it is determined that the time length is equal to or longer than L, the voice length determination unit that outputs the voice signal to the male and female voice discrimination processing unit, and the voice signal input from the voice length determination unit until the time length L is equal to or longer A voice decompression unit that repeatedly decompresses and outputs the decompressed voice signal to the male and female voice discrimination processing unit, a voice signal input from the voice length determination unit, and a voice feature amount of the voice signal input from the voice extension unit The speech signal is extracted based on the likelihood of collating the speech feature with the male acoustic model and female acoustic model using the recognition grammar corresponding to the extracted speech signal. Identify and output the gender of the speaker And a male and female voice identification processing unit.

  According to the present invention, when the time length of the input voice signal is short and less than the predetermined time length L, the voice signal is repeated, and the recognition grammar corresponding to the repetition is used. The amount normalization process can be stabilized and performed correctly, which makes it possible to accurately identify the gender of the speaker of the speech signal.

The block diagram which shows the function structure of the man and woman voice identification device which performs the 1st Example of the man and woman voice identification method by this invention. The flowchart which shows the processing flow of 1st Example of the gender voice identification method by this invention. The figure which shows the audio | voice expansion | extension example. The figure for demonstrating the effect of the audio | voice expansion | extension with respect to the normalization process of an audio | voice feature-value. The block diagram which shows the function structure of the man and woman voice identification device which performs the 2nd Example of the man and woman voice identification method by this invention. The flowchart which shows the processing flow of 2nd Example of the gender voice identification method by this invention. The figure which shows the audio | voice expansion | extension example. The flowchart which shows the processing flow of the 3rd Example of the man and woman voice identification method by this invention. The block diagram which shows the function structure of the man and woman voice identification device which performs the 4th Example of the man and woman voice identification method by this invention. The block diagram which shows the function structure of the man and woman voice identification device which performs the 5th Example of the man and woman voice identification method by this invention. The block diagram which shows the function structure of the man and woman voice identification apparatus which performs the conventional man and woman voice identification method.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

  FIG. 1 shows a functional configuration of the gender voice identification apparatus according to the first embodiment, and FIG. 2 shows a processing flow thereof.

  In this example, the gender voice identification device has a configuration in which a voice length determination unit 40 and a voice decompression unit 50 are added to the conventional gender voice discrimination device shown in FIG.

  A voice signal for identifying the gender of the speaker is input to the voice length determination unit 40 (step S1). The audio length determination unit 40 determines whether or not the time length of the input audio signal is less than the predetermined time length L (step S2), and if it is determined that the time length is less than the time length L, outputs the audio signal to the audio expansion unit 50. If it is determined that the time length is equal to or longer than L, the audio signal is output to the gender identification unit 10.

  The speech decompression unit 50 repeatedly decompresses the speech signal input from the speech length determination unit 40 until the time length becomes equal to or longer than the time length L (step S3), and outputs the decompressed speech signal to the gender voice discrimination processing unit 10.

  The time length L can be any value greater than zero. The value of the time length L may be determined by a method such as experimentally obtaining an appropriate value effective for improving the identification accuracy from a voice set of a task to which gender voice identification is applied. Here, as an example, L = 2 seconds.

  Specifically, the voice decompression in the voice decompression unit 50 is performed as follows. That is, in this example, the audio decompression unit 50 includes a buffer 51, and the buffer 51 is copied frame by frame sequentially from the first frame of the input audio signal. When the final frame of the input audio signal is reached, copying is performed again from the first frame of the input audio signal. The above processing is repeated until the frame length included in the buffer 51 becomes equal to or longer than the time length L. The copy repetitive processing here may be terminated when the frame length included in the buffer 51 becomes equal to (or exceeds) the time length L (FIG. 3 shows the decompression that is repeated and expanded in this way. An example of the audio signal is shown together with the input audio signal), or the copy is finished when the frame length included in the buffer 51 exceeds the time length L and reaches the final frame of the input audio signal. Also good.

  In general speech recognition, such speech decompression processing is not desirable because the recognition result (result of converting the speech into text) will be different from the input speech. Since it is not necessary to identify (what is spoken), such a voice decompression process can be applied.

  An audio signal is input to the gender identification processing unit 10 from the audio length determination unit 40 and the audio expansion unit 50. The voice feature quantity extraction unit 11 extracts voice feature quantities of these voice signals (step S4). The speech signal input from the speech decompression unit 50 includes speech and non-speech alternately by speech decompression, and can be repeatedly included multiple times. Therefore, it is necessary to use a recognition grammar corresponding to this repetition. The recognition grammar setting unit 12 sets the recognition grammar corresponding to each of the case where the speech signal is input from the speech length determination unit 40 and the case where the speech signal is input from the speech decompression unit 50. The recognition grammar when the speech signal is input from the speech length determination unit 40 is the recognition grammar (1) described above, and the recognition grammar when the speech signal is input from the speech decompression unit 50 is the recognition grammar (2) shown below. The Note that the recognition grammar (2) is described in an expanded form of the BNF notation, similar to the recognition grammar (1).

・ Recognition grammar (2)
$ [p] = pause;
$ [g] = garbage;
$ START = <$ p | $ g>;
However, symbol <> represents one or more repetitions, and symbol | represents parallel connection.

  The recognition grammar (2) is a grammar assuming that non-speech and speech appear alternately.

  The identification unit 13 uses the recognition grammar set by the recognition grammar setting unit 12 to collate the voice feature amount with the male acoustic model 20 and the female acoustic model 30 to obtain likelihood (step S5), and based on the obtained likelihood. Then, the gender of the speaker of the voice signal is identified (step S6). The male / female voice identification processing unit 10 outputs the result of identification in this way.

  In this example, the time length of the input audio signal is short as described above, and when the time length is less than the predetermined time length L, the time length is extended by repeating the audio signal. It is possible to increase the section in which the voice feature amount average used for identification of the gender of the speaker can be obtained.

  FIG. 4B shows this situation, and FIG. 4A shows a conventional example in which voice decompression is not performed as a comparison.

  When the normalization processing such as CMN is performed sequentially using the average of the voice feature amount for the window length N seconds (the past N seconds), the data amount that can be used for the average calculation is small in the conventional example shown in FIG. Therefore, the effect of CMN cannot be obtained sufficiently. 4A and 4B, the section indicated by the double-headed arrow indicates a window length of N seconds, and among these, the arrows indicated by solid lines all represent the average of the audio feature values for the window length of N seconds. Indicates a usable section.

In FIG. 4A, the average of the speech feature quantity for the window length N seconds cannot be used in the first section S ₁ and S ₂ , and the section in which the average of the voice feature quantity for the window length N seconds can be used is S _{3 to} S ₅ . There are 3 sections. On the other hand, in FIG. 4B in which voice expansion is performed, the section in which the average of the voice feature amount for the window length of N seconds can be used is increased to seven sections of S _{3 to} S ₉ , and thereby the effect of normalization by CMN or the like. Can be sufficiently obtained, and therefore the discrimination accuracy of male and female voices can be improved.

  Note that N is, for example, about 0.8 seconds. If N is too long, the average of a wide section is obtained, so that the normalization effect is reduced and the identification accuracy is lowered. Therefore, for example, it is not desirable to simply perform normalization processing using the speech feature values of the entire section of the speech signal, and N is set to about 0.8 seconds as described above.

  FIG. 5 shows a functional configuration of the gender voice identification apparatus according to the second embodiment, and FIG. 6 shows a processing flow thereof.

  In this example, the voice segment detection unit 60 is added to the gender voice identification apparatus of the first embodiment shown in FIG. 1, and the voice segment detection is performed with respect to the processing flow of the first embodiment shown in FIG. The process (step S11) is performed before the voice decompression process (step S3).

  An audio signal input in a general environment includes a section (non-speech section) that is not speech such as noise or silence. The same applies to a very short time audio signal, and a part of the audio signal includes a non-speech section. However, when a speech signal having a relatively short time is input, the length of the non-speech section and the length of the speech section may be approximately the same or the length of the speech section may be shorter.

  On the other hand, in the normalization of the voice feature amount, the voice section and the non-voice section are not particularly identified. For this reason, noise or the like included in the non-speech section affects the statistical analysis result of the speech feature value, and as a result, correct normalization processing may not be performed. Therefore, in a very short time speech signal including noise, etc., it is possible that an error in normalization processing due to noise or the like affects the likelihood of the male and female acoustic models and the identification accuracy is lowered. .

  The second embodiment solves this problem. The audio signal determined by the audio length determination unit 40 to be less than the predetermined time length L is input to the audio interval detection unit 60, and the audio interval detection unit 60 receives the input audio. The speech section of the signal is detected (step S11), and only the speech section is output to the speech decompression unit 50. The voice decompression unit 50 repeatedly decompresses only the input voice section until the time length becomes equal to or longer than the time length L (step S3). FIG. 7 shows a state in which a speech section is detected from an input speech signal, and the speech section is repeated to generate an expanded speech signal.

  An existing speech segment detection method can be used for speech segment detection in the speech segment detection unit 60, and for example, a speech signal segment estimation method described in Japanese Patent No. 46901079 can be used.

  In this example, the normalization process is performed only on the speech section, so that the normalization process can be stabilized and the male and female voices can be identified more accurately.

  The processing flow of Example 3 is shown in FIG. In FIG. 8, the processes of steps S12 and S13 are added to the process flow of the second embodiment shown in FIG.

  When the speech section included in the input speech signal is extremely short, there is a high possibility that the male and female voices cannot be identified with sufficient accuracy even if the speech decompression process is performed. In addition, such voice may be an erroneous voice or an input that is not a voice signal, and it may be desirable to reject it.

  In the third embodiment, this rejection is performed, and after detecting the speech section (step S11), it is determined whether the length of the speech section is less than the threshold value T (step S12) and less than the threshold value T. In this case, it is assumed that the voice section is rejected (step S13), and male and female voices are not identified. The threshold value T is a value greater than 0, and may be determined by a method such as obtaining an appropriate value experimentally. Here, as an example, T = 0.2 seconds.

  FIG. 9 shows a functional configuration of the gender voice identification apparatus according to the fourth embodiment.

  In this example, an extension time length input unit 70 is added to the gender voice identification apparatus of the first embodiment shown in FIG.

  If the audio signal to be subjected to the gender voice discrimination process is long, the time required for the gender voice discrimination process becomes long. When the demand for processing speed is strict and it is necessary to output the identification result within a certain time even if the accuracy is somewhat sacrificed, the speech decompression time length L is set to an externally set value L ′ (L ′ <L). If you can, it is convenient.

  In this example, an extension time length input unit 70 is provided to enable external setting of such extension time length, and the time required for identification (identification response time) is controlled according to the processing speed requirement. Can do.

  FIG. 10 shows a functional configuration of the gender voice identification apparatus according to the fifth embodiment.

  In this example, an extension time length calculation unit 80 is provided instead of the extension time length input unit 70 in the gender voice identification apparatus of the fourth embodiment shown in FIG.

  For example, when gender identification technology is applied to a system that recognizes an input speech signal and responds with a synthesized speech of the same gender as the speaker of the input speech signal, the input speech within approximately the same time as speech recognition. It is required to identify the gender of the signal. A time required for processing from the outside in this way is set as a request response time R.

  The decompression time length calculator 80 calculates the speech decompression time length each time from the request response time R and computer load information (for example, load average information that can be acquired from the OS) input from the outside. This is output to the voice decompression unit 50.

Here, if the computer load is W and the time length to be calculated is L ″, the time length L ″ is, for example,
L ″ = A × (R / W)
Can be calculated by: A is a constant that can be experimentally determined in advance so that a desired response time is obtained when predetermined W and L ″.

Now, in a certain system configuration, when the computer load W = 1.0, there is an experimental example that a response can be returned with a response time of 0.5 seconds if the time length L ″ = 2.0 seconds. In this case, the constant A is
A = L ″ × (W / R)
= 2.0 × (1.0 / 0.5)
= 4.0
It becomes. The constant A can be obtained in this way.

As an example of an appropriate value for the constant A, for example, 4.0, a time length L ″ is set for an input of a computer load W = 1.5 and a request response time R = 0.5.
L ″ = 4.0 × (0.5 / 1.5)
≒ 1.33 (seconds)
It can be asked.

  Using the time length L ″ obtained in this way, the voice is decompressed in the same manner as in the first embodiment. That is, the audio decompression unit 50 includes a buffer 51, and the buffer 51 is copied frame by frame sequentially from the first frame of the input audio signal. When the final frame of the input audio signal is reached, copying is performed again from the first frame of the input audio signal. The above processing is repeated until the frame length included in the buffer 51 becomes equal to or longer than the time length L ″. The copy repetitive processing here is terminated when the frame length included in the buffer 51 becomes equal to (or exceeds) the time length L ″.

  As described above, in this example, the male / female voice identification process corresponding to the required response time can be performed in consideration of the fluctuation of the computer load.

  While various embodiments have been described above, the present invention is characterized in that when an input audio signal is very short, the audio signal is repeated and expanded. On the other hand, a technique for decompressing data by combining a part of the input data with the input data is also used in the past. For example, there is a technique for interpolating data in a missing part using data in a part that is not missing when part of the transmitted data is missing. In addition, there is a technique in which when data is matched with a template, the data at the end of the data is copied outside the end of the data, and the end of the data is used as a matching target. However, none of the techniques repeatedly use the entire data for the purpose of extending the data length to an arbitrary length.

  There is also a conventional technique for performing recognition using an acoustic model. For example, voice recognition technology. However, the speech recognition technology is processing that places importance on the content of utterances that are recognized. If the input speech is repeated to extend the time length of the input speech, the utterance itself will be different.

  Accordingly, there has been no idea in the past that a technique for recognizing using an acoustic model is combined with a technique for extending the time length of an input voice by repeating the input voice.

  On the other hand, in the gender identification according to the present invention, it is not necessary to recognize the utterance content. The identification using the acoustic model is the gender of the speaker of the input voice signal, and the identification of the gender requires only the voice feature amount extracted from the input voice signal. For this reason, it is possible to improve the accuracy of male / female voice identification by performing male / female voice identification using data obtained by repeatedly expanding an input audio signal.

  The male and female voice identification device and the male and female voice identification method described above can be realized by a computer and a program installed in the computer. The program installed in the computer is decrypted by the CPU of the computer to cause the computer to execute the above-described gender voice identification method.

DESCRIPTION OF SYMBOLS 10 Male and female voice identification process part 11 Voice feature-value extraction part 12 Recognition grammar setting part 13 Identification part 20 Male voice acoustic model 30 Female voice acoustic model 40 Voice length determination part 50 Voice decompression part 60 Voice section detection part 70 Extension time length input part 80 Decompression Time length calculator

Claims (7)

A male and female voice identification method for extracting a voice feature from an input voice signal and identifying the gender of the speaker of the voice signal based on the likelihood of matching the voice feature with a male acoustic model and a female acoustic model. ,
If the time length of the audio signal is less than a predetermined time length L, the audio signal is repeatedly expanded until the time length is equal to or greater than the time length L,
A method for discriminating male and female voices, wherein the voice feature value is extracted using the expanded voice signal, and the collation and identification are performed using a recognition grammar corresponding to the repetition. The method for identifying gender voice according to claim 1,
A male / female voice identification method characterized by detecting a voice section of the voice signal and repeating the voice section to extend the voice signal when the repetition is performed. The method for identifying gender voice according to claim 2,
When the length of the detected speech segment is less than a threshold T, the speech segment is rejected and the discrimination is not performed. In the male and female voice identification method according to any one of claims 1 to 3,
The time length L can be set from the outside. In the male and female voice identification method according to any one of claims 1 to 3,
The time length L is calculated from a computer load for performing gender voice discrimination and a required response time. When it is determined whether or not the time length of the input audio signal is less than the predetermined time length L, and when it is determined that the time length is less than the time length L, the audio signal is output to the audio expansion unit and is determined to be greater than or equal to the time length L A voice length determination unit that outputs the voice signal to a gender voice identification processing unit;
The speech decompression unit that repeatedly decompresses the speech signal input from the speech length determination unit until the time length becomes equal to or greater than the time length L, and outputs the decompressed speech signal to the gender voice identification processing unit;
The speech feature is extracted from the speech signal input from the speech length determination unit and the speech signal input from the speech decompression unit, and using the recognition grammar corresponding to the speech signal from which the speech feature amount is extracted. The male and female voice models, and the male / female voice model for identifying and outputting the gender of the speaker of the voice signal from which the voice feature quantity is extracted based on the likelihood of matching. A male / female voice identification device.   A program for causing a computer to execute the gender voice identification method according to any one of claims 1 to 5.

Images