JP2005301022A - Voice feature extracting device, speaker recognizing device, program, and voice feature extracting method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a voice feature extracting device capable of analyzing individualities more in detail from the frequency spectrum of a voice without fixing the resolution in a frequency direction. <P>SOLUTION: The voice feature extracting device 4 includes a 1st analyzing means 13 of analyzing frequency components of an input voice to extract spectrum components, a logarithmic converting means 14 of logarithmically converting the extracted spectrum components, and a 2nd analyzing means 15 of taking multiple resolution analysis of the logarithmically converted spectrum to obtain a feature vector. Consequently, frequency-directional lengths of respective analysis windows are reduced together with the height of the quefrency and analysis with frequency resolution which is higher as the quefrency becomes higher is enabled to take the analysis more in detail without fixing the frequency resolution when individualities are analyzed from the frequency spectrum of the voice, thereby providing a speaker recognizing device whose speaker recognition precision is improved. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to an audio feature extraction apparatus for sequentially outputting a feature vector including personality information after dividing an input digital audio signal into frames of an appropriate length and performing window processing, and an individuality included in an audio wave The present invention relates to a speaker recognition device that recognizes a speaker using information, a program, and a voice feature extraction method.

  The automatic determination of who is the voice using the personality information included in the sound wave is called speaker recognition. Such forms of speaker recognition can be divided into speaker identification and speaker verification. Speaker identification is to determine who is the voice of the N people registered in advance. In speaker verification, an ID of who the person is is input at the same time as the input voice, and it is determined whether or not the voice is really a voice of a person corresponding to the ID. In the case of speaker identification, the speaker having the highest similarity (likelihood) is selected from a large number of registered speakers, and the speaker's voice is determined. In the case of speaker verification, if the similarity to the person's standard pattern based on the ID (likelihood for the model) is greater than a certain threshold, the person's voice is determined. Otherwise, the voice of the other person is used. It is determined that

  By the way, the performance of speaker identification is evaluated by an error rate at which a speaker other than the registered speaker is selected. Of course, as the number of registered speakers increases, it becomes more difficult, so the error rate of speaker identification increases monotonically as the number of registered speakers increases. Therefore, even when the number of registered speakers increases, it is desired to suppress the increase in the error rate of speaker identification.

  In recent years, in speaker identification, a low-order cepstrum coefficient is widely used as a feature parameter representing personality. Here, a procedure for extracting low-order cepstrum coefficients by the cepstrum method will be described with reference to FIG.

  In FIG. 7, reference numeral 501 denotes an input audio wave (digital audio signal), 502 denotes a time window processing unit that divides the audio wave into frames of an appropriate length and performs window processing such as a Hamming window, and 503 denotes a discrete Fourier. A conversion processing unit 504 is a logarithmic conversion processing unit for logarithmically converting the amplitude spectrum, 505 is an inverse discrete Fourier transform processing unit, 506 is a liftering processing unit, and 507 is an output cepstrum coefficient.

  The input audio wave 501 is divided into frames of an appropriate length (generally 20 to 30 ms) in the time window processing unit 502 and sequentially multiplied by a window such as a Hamming window. Next, an amplitude spectrum is extracted by the discrete Fourier transform unit 503, and a logarithmic amplitude spectrum is obtained by logarithmically transforming the amplitude spectrum by the logarithmic transformation processing unit 504. Generally, it is said that information indicating individuality is included in the outline information of the logarithmic amplitude spectrum envelope obtained here. In order to extract the outline information, the inverse discrete Fourier transform processing unit 505 performs inverse Fourier transform to obtain a cepstrum, and the subsequent liftering processing unit 506 removes the high-order cepstrum to obtain a low-order cepstrum coefficient ( For example, refer nonpatent literature 1).

  The inverse discrete Fourier transform is an analysis method in which the frequency resolution of the analysis window with respect to the quefrency is constant as shown in the left diagram of FIG. 8. The logarithmic amplitude spectrum is inversely discrete Fourier transformed for each analysis window, and the analysis window Find the corresponding cepstrum coefficients. A column of cepstrum coefficients obtained for each analysis window becomes a feature vector as shown in the right figure of FIG.

“Sound Information Processing” by Sadahiro Furui Morikita Publishing Co., Ltd. 25

  However, the cepstrum coefficient extracted by the conventional cepstrum method has a problem that the resolution in the frequency direction is fixed because the length of each analysis window in the frequency direction is constant. This is a factor that adversely affects speaker discrimination when information indicating individuality is biased in the frequency direction.

  The present invention provides a speech feature extraction device, a program, and a speech feature extraction method capable of performing more detailed analysis without resolution in the frequency direction being fixed when analyzing individuality from the frequency spectrum of speech. Objective.

  It is an object of the present invention to provide a speaker recognition device with improved speaker recognition accuracy.

  The speech feature extraction apparatus according to the present invention is a speech feature extraction apparatus that outputs a feature vector including personality information sequentially after dividing an input digital speech signal into frames of an appropriate length and performing window processing. Frequency analysis of the processed digital audio signal and extraction of spectral components, logarithmic conversion means for logarithmically converting the spectral components extracted by the first analyzing means, and logarithmic conversion Second analysis means for obtaining a feature vector by performing multi-resolution analysis on the logarithmic spectrum logarithmically converted by the means.

  Therefore, the length in the frequency direction of each analysis window is reduced together with the height of the quefrency to be analyzed, and analysis with high resolution in the frequency direction can be performed.

  According to the speech feature extraction device, the program, and the speech feature extraction method of the present invention, when analyzing personality from the frequency spectrum of speech, the resolution in the frequency direction is not fixed, and more detailed analysis can be performed.

  According to the speaker recognition device of the present invention, when analyzing personality from the frequency spectrum of speech, the resolution in the frequency direction is not fixed and more detailed analysis can be performed, so that the speaker recognition accuracy is improved. A speaker recognition device can be provided.

  An embodiment of the present invention will be described with reference to FIGS.

  FIG. 1 is a block diagram showing the configuration of the speaker recognition apparatus 100 of the present embodiment. As shown in FIG. 1, the speaker recognition apparatus 100 includes a microphone 1, a low-pass filter 2, an A / D conversion unit 3, a feature vector generation unit 4, a speaker selection unit 5, a speaker model generation unit 6, and a storage. It consists of part 7.

  The microphone 1 converts input sound into an electrical analog signal. The low-pass filter 2 cuts and outputs a frequency equal to or higher than a predetermined frequency from the input analog signal. The A / D converter 3 converts the input analog signal into a digital signal with a predetermined sampling frequency and the number of quantization bits. As described above, the microphone 1, the low-pass filter 2, and the A / D converter 3 constitute an audio input unit for inputting audio.

  The feature vector generation unit 4 functions as an audio feature extraction device, extracts personality feature information from an input digital signal, and sequentially outputs feature vectors that are feature data including personality information. .

  The speaker model generation unit 6 (model creation unit) creates a speaker model (personality feature model) from the feature vector generated by the feature vector generation unit 4, and the storage unit 7 (registration unit) A speaker model (for example, a code book) created by the speaker model generation unit 6 is registered.

  The speaker selection unit 5 (speaker selection means) obtains the highest similarity (likelihood) from the feature vector generated by the feature vector generation unit 4 and a speaker model (for example, a code book) registered in the storage unit 7 in advance. ) Is selected, and the selected speaker recognition result is output.

  Here, various processing units of the feature vector generation unit 4 that exhibits the characteristic functions of the present embodiment will be described with reference to FIG. As shown in FIG. 2, 11 is an input audio wave (digital audio signal), 12 is a time window processing unit that divides the audio wave into frames of an appropriate length and performs window processing such as a Hamming window, 13 Is a discrete Fourier transform processing unit (first analysis unit) that performs frequency analysis of input speech and extracts spectral components, 14 is a logarithmic transformation processing unit (logarithmic transformation unit) that performs logarithmic transformation of the amplitude spectrum, and 15 is a spectrum obtained by wavelet transformation. An MRA processing unit (second analysis means) 16 obtains a feature vector by performing multi-resolution analysis (MRA) on the components, and 16 is a feature vector (multi-resolution parameter) output by the above processing.

  As shown in the left diagram of FIG. 3, the multi-resolution analysis performed by the feature vector generation unit 4 is performed as the quefrency becomes higher because the length in the frequency direction of each analysis window is reduced with the height of the quefrency. Analysis with high frequency resolution can be performed. The feature vector generation unit 4 outputs such a feature vector (multi-resolution parameter) as shown in the right diagram of FIG. 3 by performing such an analysis.

  Note that the feature vector generation unit 4 of the present embodiment may also include a conventional analysis processing unit (third analysis unit) that extracts cepstrum coefficients as shown in FIG. By combining the conventional analysis processing unit (third analysis unit) that extracts the cepstrum coefficient, the feature vector output from the feature vector generation unit 4 is a low-order cepstrum coefficient as shown in FIG. It becomes a multidimensional vector combined with multi-resolution parameters (integration means).

  Further, when the feature vector is generated by integrating the low-order cepstrum coefficient and the multi-resolution parameter, the feature vector is not limited to adding and integrating the low-order cepstrum coefficient and the multi-resolution parameter as shown in FIG. Alternatively, it may be alternatively integrated for each cefency band which is a variable of cepstrum (integration means). By doing in this way, the effect equivalent to having employ | adopted the optimal analysis window for every quefrency zone | band can be acquired, and it becomes possible to implement a more ideal analysis.

  Next, the flow of registration processing of the speaker recognition apparatus 100 in the present embodiment will be described. The sound that has been manpowered by the microphone 1 is output as an electrical analog signal. The input sound output as an analog signal is cut by the low-pass filter 2 at a frequency that is 1/2 or more of the sampling frequency (for example, 12 kHz). Thereafter, the input sound is sampled at the sampling frequency by the A / D converter 3 and converted into a digital signal.

  The input speech converted into a digital signal by the A / D conversion unit 3 is input to the feature vector generation unit 4, and feature data including personality information extracted by speech analysis is output as a feature vector (multi-resolution parameter). Is done.

  The feature vector (multi-resolution parameter) output from the feature vector generation unit 4 is input to the speaker model generation unit 6, and a speaker model (for example, a code book) is created in the speaker model generation unit 6. A speaker model (for example, a code book) created by the model generation unit 6 is registered in the storage unit 7.

  Next, the flow of speaker recognition processing of the speaker recognition device 100 in the present embodiment will be described. The sound that has been manpowered by the microphone 1 is output as an electrical analog signal. The input sound output as an analog signal is cut by the low-pass filter 2 at a frequency that is 1/2 or more of the sampling frequency (for example, 12 kHz). Thereafter, the input sound is sampled at the sampling frequency by the A / D converter 3 and converted into a digital signal.

  The input speech converted into a digital signal by the A / D conversion unit 3 is input to the feature vector generation unit 4, and feature data including personality information extracted by speech analysis is output as a feature vector (multi-resolution parameter). Is done.

  The feature vector (multi-resolution parameter) output from the feature vector generation unit 4 is input to the speaker selection unit 5 and the highest similarity (likelihood) from a speaker model (for example, a code book) registered in the storage unit 7 in advance. A speaker with a high degree) is selected, and the selected speaker recognition result is output.

  As described above, according to the present embodiment, the length in the frequency direction of each analysis window is reduced with the height of the quefrency, and it becomes possible to perform an analysis with a high frequency resolution as the quefrency becomes higher. When analyzing personality from the frequency spectrum of speech, the frequency resolution is not fixed, and more detailed analysis can be performed. Therefore, the speaker recognition device 100 with improved speaker recognition accuracy can be provided.

  The present invention is not limited to the specific hardware configuration shown in the above-described embodiment, and can be realized by software. FIG. 6 is a block diagram illustrating a configuration example of the speaker recognition device 100 when the present invention is implemented by software. The speaker recognition device 100 includes a CPU 101 that centrally controls each unit of the speaker recognition device 100. The CPU 101 includes a ROM that stores a BIOS and a RAM that stores various data in a rewritable manner. A memory 102 is connected to the bus and constitutes a microcomputer. The CPU 101 includes a HDD (Hard Disk Drive) 103, a CD-ROM drive 105 that reads a CD (Compact Disc) -ROM 104 that is a computer-readable storage medium, and communication between the speaker recognition device 100 and the Internet. A communication device 106 that manages the above, a keyboard 107, a display device 108 such as a CRT or LCD, and the microphone 1 are connected by bus via an I / O (not shown).

  A computer-readable storage medium such as the CD-ROM 104 stores a program for realizing the voice feature extraction function of the present invention. By installing this program in the speaker recognition device 100, the CPU 101 stores the voice of the present invention. The feature extraction function can be executed. Also, the sound input from the microphone 1 is temporarily stored in the HDD 103 or the like. When the program is started, the voice data temporarily stored in the HDD 103 or the like is read, the voice feature extraction process is executed, and the feature vector extracted by the voice feature extraction process is used for the speaker recognition process. .

  As a storage medium, not only the CD-ROM 104 but also various types of media such as various optical disks such as DVD, various magnetic disks such as various magneto-optical disks and flexible disks, and semiconductor memories can be used. Further, a program may be downloaded from a network such as the Internet and installed in the HDD 103. In this case, the storage device storing the program in the server on the transmission side is also a storage medium of the present invention. Note that the program may operate on a predetermined OS (Operating System), in which case the OS may execute a part of various processes described later, or a word processor. It may be included as part of a group of program files that constitute predetermined application software such as software or an OS.

It is a block diagram which shows the structure of the speaker recognition apparatus of one Embodiment of this invention. It is a block diagram which shows the various process parts of a feature vector production | generation part. It is a schematic diagram which shows the analysis window and feature vector (multi-resolution parameter) on a frequency quefrency plane. It is a schematic diagram which shows the modification of the feature vector output by the process in a feature vector production | generation part. It is a schematic diagram which shows another modification of the feature vector output by the process in a feature vector production | generation part. It is a block diagram which shows the structural example of the speaker recognition apparatus in the case of implement | achieving this invention by software. It is a block diagram which shows the various process parts of the conventional analysis process part which extracts a cepstrum coefficient. It is a schematic diagram which shows the analysis window and feature vector on the conventional frequency quefrency plane.

Explanation of symbols

1, 2, 3 Voice input means 4 Voice feature extraction device 5 Speaker selection means 6 Model creation means 7 Registration means 13 First analysis means 14 Logarithmic conversion means 15 Second analysis means 100 Speaker recognition device

Claims (10)

In an audio feature extraction apparatus that outputs a feature vector including personality information sequentially after dividing an input digital audio signal into frames of an appropriate length and performing window processing.
First analysis means for performing frequency analysis of the windowed digital audio signal and extracting a spectral component;
Logarithmic conversion means for logarithmically converting the spectral components extracted by the first analysis means;
A second analysis means for performing multi-resolution analysis of the logarithmic spectrum logarithmically converted by the logarithmic conversion means to obtain a feature vector;
An audio feature extraction apparatus comprising: Third analysis means for extracting a cepstrum coefficient from the spectral component extracted by the first analysis means and obtaining a feature vector;
Integration means for integrating the feature vector obtained from the second analysis means and the feature vector obtained from the third analysis means;
The speech feature extraction apparatus according to claim 1, further comprising: Alternatively, vector integration for each cefency band that is a variable of cepstrum,
The speech feature extraction apparatus according to claim 3. In a speaker recognition device that recognizes a speaker using personality information included in a sound wave,
Audio input means for inputting a digital audio signal;
4. The speech feature extraction apparatus according to claim 1, wherein after the input digital speech signal is divided into frames of an appropriate length and subjected to window processing, feature vectors including personality information are sequentially output. ,
Model creation means for creating a personality feature model from the feature vector input from the speech feature extraction device;
Registration means for registering the individuality feature model created by the model creation means;
Speaker selection means for selecting the speaker with the highest similarity (likelihood) from the individuality feature model registered by the registration means based on the feature vector output from the speech feature extraction device;
A speaker recognition device comprising: A computer-readable program for executing an audio feature extraction function for sequentially outputting a feature vector including personality information after dividing an input digital audio signal into frames of an appropriate length and performing window processing. ,
A first analysis function for performing frequency analysis on a digital audio signal subjected to window processing and extracting a spectral component;
A logarithmic conversion function for logarithmically converting the spectral components extracted by the first analysis function;
A second analysis function for multi-resolution analysis of the logarithmic spectrum logarithmically converted by the logarithmic conversion function to obtain a feature vector;
That causes the computer to execute the program. A third analysis function for extracting a cepstrum coefficient from the spectral component extracted by the first analysis function and obtaining a feature vector;
An integration function for integrating the feature vector obtained from the second analysis function and the feature vector obtained from the third analysis function;
The program according to claim 5, wherein the computer is executed. Alternatively, vector integration for each cefency band that is a variable of cepstrum,
The program according to claim 6. In an audio feature extraction method for outputting a feature vector including personality information sequentially after dividing an input digital audio signal into frames of an appropriate length and performing window processing,
A first analysis step of frequency-analyzing the windowed digital audio signal and extracting a spectral component;
A logarithmic transformation step for logarithmically transforming the spectral component extracted by the first analysis step;
A second analysis step in which a logarithmic spectrum logarithmically transformed by the logarithmic transformation step is subjected to multiresolution analysis to obtain a feature vector;
A speech feature extraction method comprising: A third analysis step of extracting a cepstrum coefficient from the spectral component extracted by the first analysis step to obtain a feature vector;
An integration step of integrating the feature vector obtained from the second analysis step and the feature vector obtained from the third analysis step;
The speech feature extraction method according to claim 8, further comprising: Alternatively, vector integration for each cefency band that is a variable of cepstrum,
The speech feature extraction method according to claim 9.

Images