JPH1055195A - Speaker characteristic discrimination method by voice recognition - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a speaker characteristic discrimination method by voice recognition capable of discriminating a characteristic of a speaker based on a voice vocalized by an unspecific speaker. SOLUTION: The voices of many speakers with different speaker's characteristics are analyzed respectively, and only an analytic parameter containing many discriminative information of a phoneme among the analytic parameters incorporated in respective analytic results is adopted, and by statistically processing all analytic results, a first voice recognition model is formed, and respective analytic results are segmented using the first voice recognition model. Then, answering relations between the analytic parameters of respective frames of respective analytic results and the model parameter of the first voice recognition model are obtained respectively, and a second voice recognition model at every speaker's characteristic is formed based on the answering relations between the analytic parameters of respective frames of respective analytic results and the model parameter of the first voice recognition model, and the parameters expressing the characteristics of the speaker among the analytic parameters of the answered frames.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

[0001] 1. Field of the Invention [0002] The present invention relates to a speaker characteristic discriminating method by speech recognition for discriminating speaker characteristics such as sex and age of the speaker by speech recognition.

[0002]

2. Description of the Related Art Methods for recognizing voices uttered by unspecified speakers have already been developed. In addition, a method of determining the sex and / or age of a speaker based on a voice uttered by an unspecified speaker is also being developed.

[0003]

SUMMARY OF THE INVENTION According to the present invention, a speaker characteristic based on speech recognition which can determine the speaker characteristics such as the sex and age of the speaker based on the voice uttered by an unspecified speaker. An object of the present invention is to provide a determination method.

[0004]

According to a first aspect of the present invention, a speaker characteristic discriminating method based on speech recognition includes a learning process and a recognition process. The learning process includes a plurality of speaker voices having different speaker characteristics. Analyzing each of the analysis results, and employing only the analysis parameters containing a large amount of phoneme identification information among the analysis parameters included in each analysis result, and statistically processing all the analysis results to obtain the first voice. Creating a recognition model, segmenting each analysis result using the first speech recognition model, and determining a correspondence between an analysis parameter of each frame of each analysis result and a model parameter of the first speech recognition model. A step of obtaining each, a correspondence relationship between an analysis parameter of each frame of each analysis result and a model parameter of the first speech recognition model, and Generating a second speech recognition model for each speaker characteristic based on the analysis parameters of the assigned frame that represent the characteristics of the speaker; Analyzing the data, segmenting the analysis result using the first speech recognition model, and determining a correspondence between an analysis parameter of each frame of the analysis result and a model parameter of the first speech recognition model, respectively. Based on the correspondence between the analysis parameters of each frame of the analysis result and the model parameters of the first speech recognition model, and the second speech recognition model for each speaker characteristic. Of the second speech recognition model for each feature of the speaker best fits any of the models. Was determined, the speaker characteristics corresponding to the model that best fit, characterized in that it comprises the step, wherein the speaker for the recognition target voice data.

[0005] A second speaker feature discrimination method based on speech recognition according to the present invention includes a learning process and a recognition process. In the learning process, speeches of a large number of speakers having different speaker characteristics are respectively subjected to FFT cepstrum analysis. Step: The first phoneme H is obtained by statistically processing all the analysis results by adopting only low-dimensional analysis parameters including a large amount of phoneme identification information among the analysis parameters included in each analysis result.
The step of creating an MM set;
Segmentation using the phoneme HMM set of the above, obtaining the correspondence between the analysis parameter of each frame of each analysis result and the phoneme HMM in the first phoneme HMM set, and analyzing each frame of each analysis result. Based on the correspondence between the parameters and the phoneme HMMs in the first phoneme HMM set, and the analysis parameters of the associated frame, based on the high-dimensional parameters expressing the features of the speaker, the characteristics of the speaker Creating a second phoneme HMM set for each
Analyzing the speech data to be recognized, segmenting the analysis result using the first phoneme HMM set, analyzing the analysis parameters of each frame of the analysis result and the first
Determining the correspondence between the analysis parameters of each frame of the analysis result and the phonemes HMM in the first phoneme HMM set, and determining the correspondence between the analysis parameters of each frame of the analysis result and the phoneme HMM in the first phoneme HMM set. Based on the second phoneme HMM set, a determination is made as to which model of the second phoneme HMM set for each speaker characteristic the analysis result for the recognition target speech data best matches. And setting the feature of the speaker corresponding to the most suitable model as the feature of the speaker with respect to the recognition target speech data.

[0006] The characteristics of the speaker include age, gender, height,
It refers to items that have characteristics that can identify a speaker with good weight.

[0007]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment in which the present invention is applied to a method for determining the sex and age of a speaker will be described below.

[1] Outline of method for determining gender and age of speaker

[0009] The gender and age determination method of the speaker includes a learning process performed in advance before performing the gender and age determination;
A recognition process for determining the sex and age of the speaker based on the result of the learning process.

[1-1] Learning process

The learning process is performed as follows.

(1) Analyze the voices of many speakers with different genders and ages.

(2) The first speech is obtained by statistically processing all the analysis results by using only the analysis parameters containing a large amount of phoneme identification information among the analysis parameters included in each analysis result. Create a recognition model.

(3) Each analysis result is segmented using the first speech recognition model, and the correspondence between the analysis parameter of each frame of each analysis result and the model parameter of the speech recognition model is obtained.

(4) The correspondence between the analysis parameters of each frame of each analysis result and the model parameters of the first speech recognition model, and the sex and age of the analysis parameters of the associated frames are expressed. A second speech recognition model for each gender and age is created based on the parameters.

[1-2] Recognition processing

The recognition process is performed as follows.

(1) The speech data to be recognized is analyzed.

(2) The analysis result is segmented using the first speech recognition model, and the correspondence between the analysis parameter of each frame of the analysis result and the model parameter of the first speech recognition model is obtained.

(3) The recognition based on the correspondence between the analysis parameters of each frame of the analysis result and the model parameters of the first speech recognition model and the second speech recognition model for each gender and age. The analysis result for the target speech data determines which of the second speech recognition models for each gender and age best fits the model, and determines the gender and age corresponding to the best fitting model, The gender and age for the above-mentioned recognition target voice data are set.

[2] Specific description of a method for determining the sex and age of a speaker using HMM

Hereinafter, a method of determining the sex and age of a speaker using an HMM (Hidden Markov Model) will be specifically described.

The HMM is a stochastic model that approximates the statistical characteristics of speech by a distribution such as a Gaussian distribution. In the HMM, the correspondence between the frame and the model is the correspondence between the frame and each state. In each state, gender and age characteristics are represented by a basal distribution. Further, in the case of a subword model, a sentence or the like is recognized by connecting the subword models, and segmentation is performed according to the recognized model sequence.

In the following, a case will be described in which a Gaussian mixture HMM with a diagonal covariance matrix is used. It is also assumed that the structure of the phoneme HMM is a left-to-right type four-state three-loop structure.

[2-1] Learning Process The learning process is performed as follows.

(1) FFT cepstrum analysis of voices of many speakers having different genders and ages. Here, 12
It is assumed that eight-dimensional parameters are obtained.

(2) Of the analysis parameters included in each analysis result, only low-dimensional, for example, 1 to 16-dimensional analysis parameters containing a large amount of phoneme identification information are employed, and learning of a well-known forward backward algorithm or the like is performed. According to the rule, a first phoneme HMM set (first speech recognition model) is created.

The set of unspecified speaker phoneme HMMs is defined as A = {λ ₁ ,..., Λ _i , λ _I }. Here, i indicates a phoneme HMM number. I indicates the number of phoneme HMMs.

The i-th phoneme HMM {λ _i } is λ _i = {wi _{, s, m} , _{ai, s, q,} μi _{, s, m,} σ ² _{i, s, m} }. expressed.

Here, w _{i, s, m} , μ _{i, s, m} and σ ²
_{i, s, m} is _m in the s-th state in the i-th phoneme HMM
A branch probability, a mean vector, and a variance vector of the Gaussian distribution are respectively shown. Also, a
_{i, s, q} represents the transition probability of the i-th phoneme HMM from the s-th state to the q-th state.

(3) Next, for each analysis result (speech data O for learning) by gender and age, each analysis result is segmented using a first phoneme HMM set, and each frame of each analysis result is segmented. And the corresponding relationship between the analysis parameter and the phoneme HMM in the first phoneme HMM set. Hereinafter, this processing will be described.

The learning audio material O is defined as shown in the following Expression 1.

[0033]

(Equation 1)

O represents a learning word. o _t
This represents the feature vector at the frame number t. T
Represents the number of O frames.

The learning word O is subjected to segmentation (hereinafter referred to as Viterbi segmentation) using the Viterbi algorithm, and the word HMM corresponding to the learning word O is converted to the phoneme HM in the first phoneme HMM set.
Generated by concatenating M. Then, a correspondence Θ _c between the phoneme HMM in the obtained word HMM and each frame of the learning word O is obtained. The feature vector of the learning word O for obtaining the correspondence theta _c, feature vectors of low dimensional rich phoneme identification information is used.

The correspondence Θ _c is defined as shown in the following Expression 2.

[0037]

(Equation 2)

In Equation 2, ψ _{c, t} and θ _{c, t} are
Phoneme HMM number and phoneme H at frame number t
The numbers of the states of the MM are shown. Such processing is performed on each of the learning words O that are classified in advance by gender and age.

The phoneme HMM corresponding to the learning word O
If the sequence is ξ _c , the learning word O and the phoneme HMM sequence ＭＭ
likelihood _{P (O, Θ c | ξ} c) between the _c The formula 3,4,5
Is defined as shown below.

[0040]

(Equation 3)

[0041]

(Equation 4)

[0042]

(Equation 5)

In Equation 4, M is the number of mixture distributions in each state.

[0044] In Equation 5, D ^l is a low following the order of the o _t.

In equation (5), o _{t, d} , μ _{i, s, m, d} and _{^{σ 2 i, s, m,}} d _{is, o t, μ i, s} , m and σ ² _{i, s, m}
Respectively indicate the value of the d-th element.

The low-order degree of o _t is 0 ≦ d ≦ D
^l , and the higher order represents a range of D ^l + 1 ≦ d ≦ D.

(4) Next, the second for each gender and age
(A second speech recognition model) is obtained. That is, for each correspondence relationship theta _c obtained for each gender and age, performs the following process.

By the above Vidavi segmentation, θ
A set of frame numbers t associated with _{c, t} = i and ψc _{, t} = _s is defined as Ψc _{, i, s} . Determining a probability density function by maximum likelihood estimation from the same Ψ _{c, i,} a plurality of frames included in _s o _t. Here, a probability density function may be obtained as a mixture distribution. In this way, a second phoneme HMM set for each gender and age is determined.

It is to be noted that, instead of learning for each state, the analysis data associated with the segmentation section is obtained by using the segmentation result for each phoneme.
The second phoneme HMM set may be determined by a well-known forward backward algorithm.

Second phoneme HMM for each gender and age
(Λ ^r ) and each phoneme HMM (λ _i ^r ) are defined as shown in the following Expressions 6 and 7, respectively.

[0051]

(Equation 6)

[0052]

(Equation 7)

[2-2] Recognition processing

The recognition process is performed as follows.

(1) FFT cepstrum analysis is performed on the speech to be recognized. Then, a recognition target word O as defined in the same manner as Expression 1 is generated.

(2) For the recognition target word O, a word HMM corresponding to the recognition target word O is generated by connecting the phoneme HMMs in the first phoneme HMM set (first speech recognition model). If the utterance content is unknown, the word HMM may correspond to a phoneme string obtained as a result of recognizing the recognition target word O using the phoneme HMM. Then, the correspondence Θ _c between the phoneme HMM in the obtained word HMM and the learning word O is obtained by Viterbi segmentation. The feature vector to be recognized word O for obtaining the correspondence theta _c, feature vectors of low dimensional rich phoneme identification information is used.

(3) Phoneme HMM of the obtained correspondence Θ _c
, The likelihood P (O, Θ _c ) for each gender and age-specific second phoneme HMM set (の^r )
| Ξ _c ) ^r is determined based on equations 8, 9 and 10. Then, the gender and the age corresponding to the second phoneme HMM set having the maximum likelihood are set as the recognition results.

[0058]

(Equation 8)

[0059]

(Equation 9)

[0060]

(Equation 10)

[0061] As can be seen from Equation 10, the sex and age of the second set of phoneme HMM (lambda ^r) regarding the likelihood _{P (O, Θ c | ξ} c) in the calculation of ^r is gender And a high-dimensional parameter expressing age.

According to the above embodiment, the sex and / or age of the speaker can be determined based on the voice uttered by the unspecified speaker. The creation of the first set of phoneme HMMs and the calculation of the correspondence Θ _c are performed in a low-dimensional analysis parameter containing a large amount of phoneme identification information, for example, 1 to 16-dimensional analysis parameters among the analysis parameters included in each analysis result Only on the basis of
The calculation speed can be improved.

In the above embodiment, the case where the gender and age of the speaker are simultaneously determined based on the voice uttered by the unspecified speaker has been described. However, when only the gender of the speaker is determined, It is needless to say that the present invention can also be applied to the case where only the age is determined.

The present invention can also be applied to the case of discriminating the characteristics of the Japanese species other than the sex and age of the speaker, for example, the height and weight of the speaker.

[0065]

According to the present invention, it is possible to determine the speaker characteristics such as the sex and age of the speaker based on the voice uttered by the unspecified speaker.

Claims (3)

[Claims] 1. A learning process comprising: a learning process and a recognition process. The learning process includes the steps of analyzing voices of a large number of speakers having different characteristics of the speakers, and a phoneme among analysis parameters included in each analysis result. Creating a first speech recognition model by statistically processing all analysis results by using only analysis parameters containing a large amount of identification information, and analyzing each analysis result by using the first speech recognition model. Calculating the correspondence between the analysis parameter of each frame of each analysis result and the model parameter of the first speech recognition model, and analyzing the analysis parameter of each frame of each analysis result and the first speech. It represents the correspondence between the recognition model and the model parameters, and the speaker characteristics among the analysis parameters of the associated frame. Generating a second speech recognition model for each speaker feature based on the parameters; the recognition processing includes: analyzing the speech data to be recognized; and analyzing the analysis result using the first speech recognition model. Performing a segmentation to determine a corresponding relationship between an analysis parameter of each frame of the analysis result and a model parameter of the first speech recognition model, and an analysis parameter of each frame of the analysis result and the first speech recognition model; Based on the correspondence relationship with the model parameters and the second speech recognition model for each speaker feature, the analysis result for the recognition target speech data is obtained based on the second speech recognition model for each speaker feature. The best fit model is determined, and the characteristics of the speaker corresponding to the best fit model are recognized as described above. A speaker characteristic discriminating method based on voice recognition, comprising a step of characterizing a speaker with respect to target voice data. 2. The learning process includes a learning process and a recognition process.
Cepstrum analysis step, by adopting only low-dimensional analysis parameters including a large amount of phoneme identification information among the analysis parameters included in each analysis result, and statistically processing all analysis results,
Creating a first phoneme HMM set; segmenting each analysis result using said first phoneme HMM set; analyzing parameters of each frame of each analysis result and phoneme HMMs in said first phoneme HMM set;
And the corresponding relationship between the analysis parameter of each frame of each analysis result and the phoneme HMM in the first phoneme HMM set, and the speaker's analysis parameter among the analysis parameters of the associated frame. Generating a second phoneme HMM set for each speaker feature based on the high-dimensional parameters representing the features; and recognizing the speech data to be recognized. Segmenting using the first phoneme HMM set to determine the correspondence between the analysis parameter of each frame of the analysis result and the phoneme HMM in the first phoneme HMM set, respectively, and the analysis parameter of each frame of the analysis result And the correspondence between the phoneme HMMs in the first phoneme HMM set and the phoneme HMM for each speaker feature. Of the second phoneme HMM set for each speaker feature is determined based on the phoneme HMM set of A speaker characteristic discrimination method by voice recognition, comprising the step of setting the characteristics of the speaker corresponding to the model being performed to the characteristics of the speaker with respect to the recognition target voice data. 3. The speech recognition according to claim 1, wherein the feature of the speaker is one arbitrarily selected from age, gender, height, and weight, or any combination thereof. Speaker characteristic discrimination method by