JP2008224911A - Speaker recognition system - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a method of identifying or matching a speaker, the method being characterized in that not only power spectrum information, but also phase information which has not been used before are applied to speaker recognition. <P>SOLUTION: In the method (speaker matching) of identifying which of speeches of speaker models structured in advance on the basis of identical feature parameters a speech that somebody utters corresponds to by extracting the feature parameters representing the individuality of the speaker included in the speech or a method (speaker matching) of determining whether the speech is a speech of a certain person, an extracting method for a phase feature parameters which represents features of a speaker included in a speech waveform and has not been used before and a conventional feature parameter extracting method are used in combination to identify or match the speaker. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to speaker identification and speaker verification for personal identification using speech.

  In speaker recognition using speech, a method of extracting the personality of a speaker included in the speech uttered by a certain person and identifying who is the speech among the registered speakers (speaker identification) Then, there is a method (speaker verification) for determining whether or not the voice is of a certain person. Furthermore, there are text-dependent speaker recognition in which the utterance content is predetermined and the user is required to utter the content, and text-independent speaker recognition in which any content can be uttered. In general, the text-dependent type provides higher performance, but if a user's voice is stolen by a recording device or the like, it can be "spoofed" by playing the recording, which is dangerous for security. For this reason, there is a method for preventing “spoofing” by instructing the contents to be uttered by the user from the speaker recognition system or apparatus as needed, which is called text-designated speaker recognition. This patent covers all the individual features included in speech, such as speaker identification, speaker verification, text-independent speaker recognition, text-dependent speaker recognition, text-specific speaker recognition, etc. Is related to extracting feature parameters representing

  A typical representative characteristic parameter represents power spectrum information obtained for each short time interval of speech, and the mel frequency cepstrum coefficient is a global standard. Human hearing is sensitive to the power spectrum, but insensitive to phase information, and in the case of normal quiet environment speech, the phase difference cannot be perceived. For example, the sum of a 200 Hz sine wave and a 1100 Hz sine wave and the sum of a 100 Hz sine wave and a 1100 Hz cosine wave sound exactly the same. Therefore, for speech recognition, robust recognition can be performed by discarding phase information using only power spectrum information. Speaker recognition has been considered in exactly the same way, and phase information has not been used for speaker recognition.

  When a spectrum is obtained by Fourier transform in a short time section (for example, 32 milliseconds) of a speech waveform, a real part and an imaginary part are obtained. The magnitude of this complex number (power spectrum) for each frequency has been used as an important feature parameter. For example, if a segment of 512 points (32 milliseconds) is cut out and a discrete Fourier transform is performed with a 16 kilohertz sampling waveform, 256 complex representation spectra are obtained. If this complex number is regarded as a vector expression, the size (the square of the real part and the square of the imaginary part: power spectrum) is important, and the angle of the vector (the ratio of the real part to the imaginary part) has been ignored. It was.

  Considering the vocal organs of human voice, the sound source waveform caused by vocal cord vibration varies greatly from individual to individual, and affects not only the difference in power spectrum but also the phase difference. It has been said that the vocal tract corresponding to the acoustic tube defines the properties of phonology, which is expressed in the power spectrum. However, it is considered that different vocal tracts correspond to the same power spectrum, which appears as a phase difference.

Therefore, through experiments, it was found that speaker information exists not only in the power spectrum but also in phase information. Since this phase information has less speaker information than the power spectrum, it has been found that the phase information has an effect of interpolating the power spectrum when used together with the power spectrum.

  The problem to be solved is that not only power spectrum information but also phase information that has not been used conventionally is applied to speaker recognition.

  The problem when obtaining the phase information is that the phase changes when the cut-out section changes even in a steady speech section. In the case of cutting out 512 points, the phase of the spectrum to be obtained changes when the point of cutting out even one point is changed (however, the power spectrum does not change). Therefore, for example, this problem is solved by determining the reference frequency ω and obtaining the relative phase shift of other frequencies from this frequency. Various realization methods are conceivable for this, but as an example, normalization is performed so that the phase of ω hertz is 45 degrees (= π / 4). Accordingly, the phases of other frequencies ω ′ are also normalized. If the phase of ω is φ, and ψ is the phase ψ of another frequency ω ′, ψ is converted into ψ + ω ′ × (π / 4−φ) / ω. Thus, the relative value of the phase of ω ′ is obtained with respect to the phase of ω.

  In this way, for example, a 256-point phase sequence is obtained by 512-point discrete Fourier transform. Although it is conceivable that this is used as a feature parameter as it is, it is usually necessary to reduce the number because of the trade-off relationship such as the number of training data. In the case of a power spectrum, after conversion to logarithmic power, band pass filter processing (usually 20 to 30) is performed, then cepstrum conversion is performed, and about 12 low-orders are used ( Mel cepstrum coefficient: MFCC). The purpose of the cepstrum transformation is that the power spectrum is multiplied by the characteristics of the sound source, and the harmonics of the fundamental frequency are added, so that the power spectrum representing the vocal tract feature does not form a smooth spectral envelope. There are various methods for reducing the 256 phase feature parameters to about 12. However, as an example, the first 12 or the 12th from the 5th to the 16th may be used. . The reason why the phase parameter is not converted into a cepstrum (of course, it may be converted into a cepstrum) is that the phase parameter is not significantly affected by the harmonics of the fundamental frequency unlike the power spectrum.

  For example, 12 MFCC and 12 phase parameters are used as separate feature parameters, and the likelihood is obtained by comparing with those expressing the speaker model, and speaker identification is performed using the likelihood together. There is also a method of connecting each of the 12 feature parameters and obtaining a likelihood by matching with the speaker model as 24 feature parameters.

[Incorporation into conventional system]
It is easy to incorporate this feature parameter into a conventional speaker recognition system. This is because most conventional speaker recognition systems calculate the likelihood by extracting feature parameters from the speech waveform and comparing it with the speaker model created for each speaker, thereby maximizing the likelihood. Whether the speaker corresponds to the speaker model is determined based on whether the speaker corresponding to the model is the result (speaker identification) or whether the likelihood of matching with the speaker model exceeds a threshold value (speaker verification) ). This can be realized simply by replacing the characteristic parameter of the system with this characteristic parameter. Alternatively, the speaker recognition method based on the combined use of the likelihoods obtained from the two feature parameters can directly use the idea of Patent Document 1 below.

Note that feature parameters depend not only on the speaker but also on the phoneme. Therefore, instead of preparing one typical feature parameter for each speaker, the model should not cover variations due to various phonemes. Don't be. A typical technique is introduced in the following Patent Document 1.
JP 2005-091758 A, Speaker recognition system method

  The effectiveness was verified using the speech database for speaker recognition evaluation provided by NTT. This database is data produced by 22 male speakers and 13 female speakers over five periods of about one year. The sampling frequency is 16 kilohertz. The frame length is 25 milliseconds, and the frame period is 10 milliseconds. A speaker model was created with five sentences spoken during this initial period. A Gaussian mixture model was used as a model (GMM). There are three types of utterance speeds: normal, fast, and slow, but the speaker model used five sentences uttered at normal speed. The experiment is performed for each sentence, and the total number of test sample sentences is 700 utterance sentences. The identification rate when using MFCC, which is a global standard feature parameter, was 95.7%. On the other hand, when the first 12 phase parameters were used, it was 41.0%. When speaker identification was performed using the likelihood obtained by combining both of these likelihoods, it was improved to 97.6%.

  A function for extracting feature parameters of phase information is added to the feature extraction unit of the conventional speaker recognition system. Accordingly, speaker model creation, likelihood calculation, and speaker determination unit are improved and realized.

  FIG. 1 shows an embodiment of a speaker recognition system. First, in the training phase, a speaker model is created for each speaker from a set of speech data uttered by the speaker in offline processing.

  First, in FIG. 1, the voice analysis unit 11 captures a voice waveform into the system. This corresponds to discrete Fourier transform and linear prediction analysis (LPC) for speech waveforms. For example, it is converted into discrete sample signal time-series data obtained by sampling 16 kilohertz and AD-converting 16 bits per sample. This is multiplied by the analysis window to cut out the speech waveform. This time length is called a frame. The next frame is cut out while shifting so that adjacent sections overlap. This interval of shifting is called a frame period. For example, 256 points are cut out by applying a Hamming window. If this is subjected to discrete Fourier transform, 128 complex representation spectra can be obtained. In addition to this, a spectrum is obtained by linear prediction analysis (LPC analysis) or the like. This is calculated every 10 milliseconds of the frame period.

  In FIG. 1, the feature extraction unit 12 extracts feature parameters based on the voice analysis result. In addition to MFCC, which is a typical representative feature parameter, phase feature parameters are extracted. Usually, the complex spectrum is converted into a power spectrum which is the sum of the square of the real part and the square of the imaginary part. The logarithm is taken and the frequency axis is converted to a mel scale, and then it is passed through several tens of band-pass filter groups and subjected to discrete cosine transform, and about 12 low-order parameters are used as characteristic parameters (MFCC). On the other hand, in order to obtain phase information, an angle is obtained from the ratio of the real part to the imaginary part (phase). Since this phase varies depending on the cutout position, the phases of other frequencies are normalized to relative values so that a certain reference frequency ω becomes π / 4. Among these 128 values, an appropriate value of about 12 is used in consideration of the trade-off between training data amount and recognition accuracy. Various methods can be used to determine which 12 are used. For example, the first 12 are used. This corresponds to a frequency band of 60 to 720 hertz. These feature parameters are calculated for each frame period and become time series parameters.

  In FIG. 1, the speaker model creation unit 13 creates a speaker model for each speaker using the feature parameters obtained by the feature extraction unit 12. Typical models include GMM and HMM. As the speaker model, a Gaussian mixture distribution (GMM) or a hidden Markov model (HMM) is usually used. Although the normalization phase varies depending on the speaker, even the same speaker varies depending on the phoneme. Therefore, this feature pattern set is expressed by a mixture of a plurality of distributions. The feature parameters are, for example, 12 MFCC and 12 normalized phases are handled separately, and a speaker model is created for each of them, and both are connected to create a speaker model as 24 parameters. A way to do this is conceivable. In the test phase for recognizing the speaker from the actual utterance sample, the speech analysis unit 11 and the feature extraction unit 12 are the same as in the training phase. However, this face is online real-time processing.

  In FIG. 1, the likelihood calculation unit 14 calculates the likelihood using the speaker model created by the speaker model creation unit 13 for the feature parameter time series of the test data.

In FIG. 1, the speaker recognition determination unit 15 identifies a speaker or collates the speaker based on the likelihood obtained by the likelihood calculation unit 14. In speaker identification, the likelihoods of the speaker models are compared, and the speaker corresponding to the speaker model with the highest likelihood is output as the identification result. On the other hand, in the case of speaker verification, a predetermined threshold value is compared with the likelihood obtained by the likelihood calculating unit 14, and if the likelihood is larger, it is determined that the speaker corresponds to this speaker model. judge. If the likelihood is less than the threshold, it is determined that the test sample speaker is not a speaker model speaker. At this time, when there are two likelihoods obtained by the likelihood calculating unit 14 such as the likelihood by the MFCC and the likelihood by the normalized phase, the weighted linear sum of these likelihoods is regarded as the likelihood.

Wide range of applications such as building entrance management, personal management that replaces PC passwords, provision of services suitable for individuals by personal identification within groups such as homes and hospital rooms, and automatic assignment of speakers when creating conference minutes It is possible.

It is a block diagram concerning the example of a speaker recognition system.

Explanation of symbols

DESCRIPTION OF SYMBOLS 11 ... Speech analysis part 12 ... Feature extraction part 13 ... Speaker model preparation part 14 ... Likelihood calculation part 15 ... Speaker recognition determination part MFCC ... Mel-Frequency Cepstrum Coefficient (Mel frequency cepstrum coefficient)
GMM ... Gaussian Mixture Model
HMM… Hidden Markov Model

Claims (1)

A method of extracting the personality of a speaker included in the voice uttered by a certain person and identifying who is the voice among the registered speakers (speaker identification), or whether it is the voice of a certain person In the method (speaker verification) for determining the speaker, a phase feature parameter extraction method that has not been used in the past and representing speaker characteristics included in the speech waveform, and speaker identification or speaker verification using the same are performed. Technique.

Images