JPH1185181A - Generating method of voice model and method and device for speaker recognition using the model - Google Patents

Abstract

PROBLEM TO BE SOLVED: To generate the voice model, on which discriminating and collating capabilities are maintained at a high level against the speaker's voice that varies depending on the uttering time, and to realize such a speaker recognition that a high recognition performance is maintained while utilizing the model. SOLUTION: The feature parameter string, which is extracted from latest inputted voices in a feature parameter extracting section 1, and the feature parameter string, which is previously stored in a feature parameter storage section 2, are inputted to a model generating section 3. In the section 3, a parameter θ<(t)> of a hidden Markov model(HMM) is defined as the one obtd. by such a manner that a parameter θ' of the HMM, that is estimated from time independent voice data components, is transformed by a model transformation function G<(t)> , that represents the variation being dependent on time t. Then, the parameter θ' and the function G<(t)> of each time are estimated based on the inputted feature parameter string at each time. Then, the parameter θ' is stored in a model storage section 4 and used for a speaker recognition when recognition voice data are inputted.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for generating a speech model based on a characteristic parameter sequence representing a speaker's personality extracted from input speech, a speaker recognition method using the speech model, and an apparatus therefor.

[0002]

2. Description of the Related Art Conventionally, as a technique used for recognizing who a visitor is from the sound of an intercom, or identifying the same person as a personal identification number based on the input sound, etc. , Extracting a feature parameter sequence representing the speaker's personality from the input speech,
By calculating the similarity with the model of the voice modeled by the feature parameter sequence registered for each speaker in advance,
Recognition of a speaker who uttered an input voice has been studied. In such speaker recognition, cepstrum, pitch, etc. are used as feature parameters representing the individuality of a speaker, and a speech model registered for each speaker is referred to as “hidden Marko”.
v model ”(Hidden Markov Model; hereinafter, referred to as“ HMM ”), a feature parameter sequence included in speech such as sentences uttered by the speaker, and a registered HM of each speaker.
A method of making a determination based on the similarity to M is often used.

Here, with regard to a speech model used in such speaker recognition, an initial model is often created from a small amount of data uttered by each speaker at one time in an actual system in consideration of the utterance load of the speaker. . However, the voice of the speaker is especially 2
It fluctuates greatly in units of up to three months (for example, see the document "Tadahiro Furui:" Individuality information included in speech waves ", Dissertation of the University of Tokyo, 1978), and the direction of the fluctuation is not constant. However, high recognition performance cannot be maintained with the initial model.

[0004] For this reason, conventionally, each speaker is made to utter periodically to accumulate the data,
A method of regenerating the HMM of each speaker using the plurality of data has been studied.

[0005]

However, a plurality of data accumulated by periodically uttering each speaker,
Includes various vocal variations. For this reason, the above-mentioned conventional method has a problem that the variance of the distribution of the feature parameter sequence of the regenerated HMM is widened, and the discriminating ability and the matching ability are reduced.

[0006] The present invention has been made in view of such circumstances, and the present invention relates to a speaker's voice that fluctuates depending on the utterance time.
An object of the present invention is to provide a method for generating a speech model that can maintain a high level of discrimination ability and matching ability. or,
It is an object of the present invention to provide a speaker recognition method and apparatus capable of maintaining high recognition performance using such a speech model.

[0007]

According to the present invention, it is assumed that (1) data of voices uttered by a speaker at a plurality of periods are obtained from a plurality of populations corresponding to each period. (2) The plurality of populations commonly have independent speaker characteristics at different times. Thus, it is considered that the data of the voice uttered by each speaker at each time is divided into a variable component depending on the utterance time and a component independent at each time for each speaker.

Assume that a parameter of the HMM estimated from data of a voice uttered by a certain speaker at time t is θ ^(t) .
At this time, the parameter θ ^(t) is the parameter θ of the HMM estimated from the independent voice data component at the time of the speaker.
＾ is a model conversion function G representing a variation depending on the time t.
^If it is converted by ^(t) , it can be expressed as shown in Equation 1 below.

(Equation 1) In addition, the symbol "$" attached above the character in the formula corresponds to the symbol "$" added after the character in the specification (the same applies to the following formulas).

The time independent parameter θ 時期 in the above equation
Is a form in which the time-dependent fluctuation component is removed from the parameter θ ^(t), so that its variance can be expected to be small. Therefore, the present invention provides independent H
The speech model is generated by estimating the parameters of the MM using the feature parameter sequence extracted from the input speech.

Here, the above-described HMM parameter θ ＾ independent of the timing is expressed by the following equation (1).
^Assuming that ^(t) is expressed, a parameter θ common to all of the timings and a set G _set (= (G ⁽¹⁾ ,
G ⁽²⁾ ,..., G ^(T) )), which can be estimated by obtaining the combination that maximizes the likelihood for the audio data O ^{(t) at} each time.

That is, as shown in the following equation 2, audio data O ^(t) (t = 1, 2,..., T) of all available periods is used.
A set of parameters θ ＾ that maximizes the likelihood for speech data and a model conversion function G _set ＾ = (G ＾ ⁽¹⁾ , G
＾ ⁽²⁾ ,..., G ＾ ^(T) ) at the same time, and θ ＾ thereof is used as a parameter of the speaker's speech model.

(Equation 2) Note that “L (O ^(t) ; G ^(t) , θ)” on the right side of the equation
Represents the likelihood function of the voice model G ^(t) (θ) with respect to the voice data O ^(t) , and “arg” indicates the right side of the whole time period (t = 1 to T)
Is meant to be a set of maximizing the direct product of the likelihood of the speech model _{(θ ^, G s et ^} ) for each sound data. And the invention described in each claim of this application based on the above theory is as follows.

According to a first aspect of the present invention, there is provided a voice model generating method for extracting a characteristic parameter representing a personality of a speaker from voice and generating a voice model of the speaker based on the characteristic parameter. The feature parameters are extracted from the voice uttered in the above, the parameters extracted from the time-independent components of the voice are estimated from the feature parameters at each time, and the speech that utters the voice based on the parameters is estimated. It is characterized by generating a voice model of a person.

According to a second aspect of the present invention, there is provided a voice model generating method for extracting a characteristic parameter representing a personality of a speaker from voice and generating a voice model of the speaker based on the characteristic parameter. The characteristic parameters are extracted from the voice uttered at the same time, and from the characteristic parameters at each time, the parameters common to each time extracted from the time-independent component of the voice, and the variation depending on each time. It is characterized by estimating a conversion factor to be converted correspondingly and generating a voice model of the speaker who uttered the voice based on the parameter common to each time.

According to a third aspect of the present invention, in the method of generating a speech model according to the second aspect, for each set of the estimated common parameter and the conversion factor at each time, the common parameter at each time is converted by the conversion factor. Calculating the likelihood of the characteristic parameter of each period estimated by conversion to the voice, and generating the voice model based on a parameter common to each time included in the set in which the likelihood is maximized. And

According to a fourth aspect of the present invention, for each speaker, each voice model is generated in advance using the voice model generating method according to any one of the first to third aspects, and the unspecified speech is generated. A feature parameter is extracted from the voice uttered by the person, and a similarity between the feature parameter and the voice model is obtained;
It is characterized in that the unspecified speaker is recognized based on the similarity.

According to a fifth aspect of the present invention, there is provided a feature parameter extracting means for extracting a feature parameter representing a speaker's personality from an input voice, and a feature parameter extracted from the input voice for a plurality of periods. Model generation means for estimating a parameter extracted from a component independent of time and generating a voice model of a speaker who uttered the input voice based on the parameter, and a model storage for storing the generated voice model Means for recognizing the unspecified speaker based on the similarity between the feature parameter extracted from the input voice uttered by the unspecified speaker and the accumulated speech model, and recognizing the unspecified speaker. It is characterized by having.

According to a sixth aspect of the present invention, a feature parameter extracting means, a model storing means and a recognizing means according to the fifth aspect, and a feature parameter of each time are extracted from a time-independent component of the input speech. The parameters common to each time period and the conversion factors for converting the parameters with a variation corresponding to each time period are estimated, and a voice model of the speaker who uttered the input voice is generated based on the parameters common to each time period. And a model generating means.

According to a seventh aspect of the present invention, in the speaker recognition apparatus according to the sixth aspect, the model generating means includes a set of a parameter and a conversion factor which are estimated at each time and which are common to each time. The likelihood for the input voice of the feature parameter of each period estimated by converting the parameter by the conversion factor is determined, and the likelihood of the voice model is determined based on the common parameter for each period included in the set in which the likelihood is maximized. The generation is performed.

The invention according to claim 8 is a step of extracting characteristic parameters each representing a speaker's personality from input voices uttered at a plurality of periods, and extracting the characteristic parameters of the input voice from the extracted characteristic parameters at each period. A program for causing a predetermined arithmetic device to execute a process of estimating a parameter extracted from a component independent of the time and a process of generating a voice model of a speaker who uttered the voice based on the parameter are executed. It is a recording medium on which recording is performed.

According to a ninth aspect of the present invention, there is provided a method of extracting a characteristic parameter representing a speaker's personality from an input voice, and extracting a characteristic parameter extracted from a plurality of input voices in a time of the input voice. Estimating parameters extracted from the independent components and generating a speech model of the speaker who uttered the input speech based on the parameters; accumulating the generated speech models; Determining a similarity between the feature parameter extracted from the input speech uttered by the user and the stored speech model, and recognizing the unspecified speaker based on the similarity. Is a recording medium on which the program is recorded.

[0021]

BEST MODE FOR CARRYING OUT THE INVENTION

<Configuration> An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing a configuration of a speaker recognition device according to one embodiment of the present invention.

In FIG. 1, reference numeral 1 denotes a feature parameter extraction unit for converting voice data into feature parameters such as cepstrum and pitch. The feature parameter extraction unit 2 sequentially converts input voice data for a certain period of time into feature parameters. Then, by sequentially supplying them to the model generation unit 3 or the similarity calculation unit 5, a feature parameter sequence obtained by converting the input voice data into a time series is input to each of these units. The feature parameter storage unit 2 is a storage unit that stores the feature parameters supplied from the feature parameter extraction unit 1, and stores a time series (feature parameter sequence) of the previously stored feature parameters.
Is input to the model generation unit 3 as needed (when generating a voice model; details will be described later).

The model generating section 3 is based on the feature parameter string input from the feature parameter extracting section 1 and the feature parameter string input from the feature parameter storing section 2.
A voice model independent of vocal fluctuations at each time is generated, and the voice model (a parameter representing the model) is output to the model storage unit 4. The generation of the voice model by the model generation unit 3 will be described in detail in the operation description described later.
The model storage unit 4 stores the speech model output from the model generation unit 3 and, if necessary (when performing speaker recognition.
The details will be described later.
Supply to

The similarity calculating section 5 includes a feature parameter sequence supplied from the feature parameter extracting section 1 and a model storing section 4.
And calculates the degree of similarity with the speech model stored in the speaker recognition determining unit 6.
Output to The speaker recognition determination unit 6 reads a predetermined threshold value from the threshold value accumulation unit 7 and compares the read threshold value with the similarity input from the similarity calculation unit 5 to determine a speaker. Here, the predetermined threshold value is a value indicating a variation range of the similarity that can be regarded as a speaker's voice corresponding to the speech model used in the calculation of the similarity. , The threshold value for each speaker is stored in advance.

<Operation> (1) Generation of Speech Model Next, the operation of the above configuration will be described. First, an operation of generating a speaker voice model (operation at a stage of registering a speaker) will be described. In this case, the data of the voice uttered periodically by the speaker to be registered is sequentially input to the feature parameter extracting unit 1. As a result, the feature parameter extraction unit 1 converts the speech data at each utterance time into a time series of feature parameters, and supplies the time series to the feature parameter storage unit 2 and the model generation unit 3 sequentially. Is generated.

FIG. 2 shows a procedure for generating a voice model in the model generating section 3. Here, the HMM satisfying the above equation (1)
Is estimated by a method corresponding to the above equation (2). Then, as an example of the estimation method, a document “T.Anastasakos, J. McDonough, R. Schwartz, and
J. Makhoul, "A Compact Model for Speaker-Adaptive T
Raining ", Proc.ICSLP, pp.1137-1140, 1996.", using the speaker adaptive training method (SAT method), and using a weighted addition of multiple Gaussian distributions for speech models independent of the utterance time. Estimate a parameter θ ＾ to be represented by the generated HMM.

A feature parameter sequence obtained by converting the latest input speech data is input from the feature parameter extracting unit 1 to the model generating unit 3, and a previously extracted feature parameter sequence at each time is stored in the feature parameter storing unit 2. (Step S1). In the following, the time at which the latest input voice is uttered is T, and the respective times at which the voice from which the characteristic parameter sequence input from the characteristic parameter storage unit 2 is extracted are uttered are 1 to T-1. .

Thereafter, the model generator 3 estimates a model conversion function at each time, based on the input characteristic parameters at each time (step S2). Here, it has been confirmed that the average vector of the Gaussian distribution of the HMM has a bias component that depends on each period on average. Based on this, in estimating the model conversion function here, H
With respect to the state j of the MM and the average vector μ _jk of the Gaussian distribution k, a model conversion function G ^(t) at the time t is defined as shown in the following Expression 3.

(Equation 3)

In Equation 3, b ^(t) is a bias component depending on time t. That is, the model conversion function G ^(t) is obtained by adding the bias component b ^(t) to obtain an average vector μ _{jk の} of a Gaussian distribution k composed of components independent of time.
A Gaussian distribution mean vector μ ^(t) _jk representing the characteristic parameter sequence at time t is converted (however, since the object to be converted is represented by a vector, G ^(t)
Is represented by a matrix. ). Therefore, estimating the model conversion function at each time in step S2 specifically corresponds to ^obtaining the bias component b ^(t) at each time.

In this embodiment, the reference "A. Sankar and C.-
H. Lee, "Robust speech recognitionbased on stochasti
c matching ", Proc. ICASSP, pp. I-121-124, 1995."
The bias component at each time is determined according to the following equation 4 so that the direct product of the right side is locally maximum (the bias component determined thereby is represented by b ^(t) ）).

(Equation 4) Wherein, N _t represents the data length in time t.

Here, when the stochastic matching method is used, it is necessary to obtain an initial model from the input data. Equation 4 is a form using a variance value or the like in the initial model. I have. Therefore, the model generation unit 3 first uses the conventional HMH method with the conventional method of estimating the parameters of the HMM by the maximum likelihood using all the input data (the characteristic parameter sequence at all times from t = 1 to T).
Estimate the initial model represented by M.

Then, in the initial model, the probability γ ^(t) _jk (n) that the observation vector o ^{(t )} (n) appears in the state j and the Gaussian distribution k, and the average vector of the Gaussian distribution k representing the initial model Using the μ _jk and the variance σ _jk , a bias component b ^(t)の at each time is obtained according to the above ^{equation (} 4 ⁾ .
The observation vector o ^(t) (n) is a vector having a time series of feature parameters extracted from the audio data at the time t as components.

Next, the average of each Gaussian distribution for generating the HMM is estimated based on the model conversion function estimated in step S2 (step S3). Specifically, using the bias component b ^(t)求 め obtained as described above, the average vector μ _{jk 状態 of} the state j and the (uncorrelated) Gaussian distribution k is calculated by the following equation (5).

(Equation 5) Γ ^(t) _jk (n) in this equation is the observation vector o ^(t) (n) in the state j and Gaussian distribution k in the HMM.
Is the probability that appears.

Subsequently, each gas estimated in the above step S3 is
Estimate the variance of each Gaussian distribution based on the mean of the male distribution
(Step S4). Specifically, in step S3
Average vector μ of Gaussian distribution k found_jkUse ＾
The variance vector σ is given by Equation 6. _jkEach element σ of ＾_jkl＾ (l
Is an identifier representing a vector element. ) Calculate the minutes
Scatter vector σ_jkAsk for ＾.

(Equation 6) _Where μ _jkl ^(t) _で is the l-th element of the mean vector μ _jk ^(t) ＾. Where the mean vector μ _jk ^(t) 、 is
An average vector (ie, μ _jk ^(t) adapted to the data at time t obtained by converting the average vector μ _{jk の} of the Gaussian distribution k determined in step S3 by the model conversion function G ^(t) determined in step S2 ⁾ ^ = a ^{_{μ j k + b (t)}} ).

The model generating unit 3 uses the average vector μ _jk ＾ and the variance vector σ _{jk 得} obtained as described above as parameters of the HMM of the speaker (corresponding to the above parameter θ ＾). 4 to terminate the generation of the speaker's voice model (step S5). Thereby, the model storage unit 4 stores the output parameters as a voice model of the speaker. At this time, each Gaussian distribution represented by each element of the mean vector μ _jk ＾ and the variance vector σ _jkを is appropriately weighted.

The above series of processing is performed for each speaker to be registered, and a voice model of each speaker is generated and stored in the model storage unit 4. In addition, even when new audio data is input after a lapse of time, an audio model can be generated in the same manner as described above by inputting the new audio data to the feature parameter extracting unit 1.

(2) Speaker Recognition Next, an operation of recognizing a speaker using the speech model generated as described above will be described. In the step of recognizing the speaker, the voice data of the target speaker is input to the feature parameter extracting unit 1 as voice data for speaker recognition. Then, the feature parameter extraction unit 1 converts the audio data into a time series of feature parameters and supplies the time series to the similarity calculation unit 5.

The similarity calculator 5 calculates the similarity between the input feature parameter sequence and the speech model stored in the model storage unit 4. In the present embodiment, the determination of the identity with a specific speaker (speaker verification) is performed, and the I
It is assumed that a speech model for calculating the similarity is specified by, for example, inputting D. Thereby, the similarity calculating unit 5
Reads out the specified voice model from the model storage unit 4, calculates the similarity between the input feature parameter sequence and the read voice model, and sends the calculated similarity value to the speaker recognition determination unit 6. Output.

Subsequently, the speaker recognition judging section 6 reads a threshold value indicating a range of similarity variation that can be regarded as a speaker's voice corresponding to the ID from the threshold value accumulating section 7, and reads out the similarity calculating section 5 Is compared with the similarity input from. Accordingly, if the value of the similarity is greater than the read threshold value, the voice of the target speaker is determined to be the voice of the speaker corresponding to the ID, and if the value is lower, the voice of the target speaker is determined. Is the above I
D is determined to be the voice of another person other than the speaker corresponding to D,
The result of the determination is output.

The generation of the speech model and the speaker recognition described above may be recorded on a recording medium as a program for causing a predetermined arithmetic unit to execute the above-described operation procedures. By causing the arithmetic unit to read the recording medium, the above-described generation of the speech model and speaker recognition may be performed.

[0041]

Next, an embodiment of the above-described embodiment based on an experiment in which a speech model was actually generated using speech data and speaker recognition was performed will be described. In the experiment mode, a text-independent speaker verification experiment will be described as an example.

Voice data The voice data to be used includes seven males (at periods T1, T2,..., T7) for approximately 16 months.
7), sentence data uttered in a quiet room is used. Here, of the 20 males, 10 are registered speakers and the other 10 are impostors. The sentence text of the sentence data is extracted from 503 sentences of the ATR continuous speech text, and the number of uttered sentences at each time is set to 5 for generation of a speech model and 3 for a test of speech recognition. The texts of the five sentences for generating the speech model are different for each speaker at each time, and the length of one sentence is about 4 seconds on average. On the other hand, the text of the three sentences for the test is different from the text of the five sentences, but is the same for all speakers and all times.

Characteristic Parameter The characteristic parameter is a cepstrum, which is a characteristic amount conventionally used, and is used by converting audio data into a time series of fine cepstrum every time. This cepstrum has a sampling frequency of 12 kHz, a frame length of 32 ms, a frame period of 8 ms, and LPC analysis (Linear Predictive Co.).
ding; linear prediction analysis)

Speech Model As a speech model, a continuous HMM having one state and 16 uncorrelated Gaussian mixture is used. In the generation of the voice model, sequential learning is performed by initially using the voice data of the five sentences uttered at the time T1, and thereafter using the voice data of the five sentences uttered at each time sequentially. It shall be. The expression of the speaker's HMM by weighted addition of 16 Gaussian distributions is described in, for example, the document "Tomoko Matsui, Sadahiro Furui:" VQ, Comparative Study of Text-Independent Speaker Recognition Method Using Discrete / Continuous HMM "" , IEICE Symposium, SP91-89, 1991. "

Speaker Recognition In the speaker recognition test, the above three sentences uttered at the time following the time at which the sentence used in the generation of the speech model was uttered are used. At this time, the first one second of each of the three sentences is actually used for the determination.

FIG. 3 shows the utterance timing of the voice data used in the learning (generation of the voice model) and the test under the above conditions. In this figure, in the upper cases X, A, B, C, D, and E, a voice model is generated using the voice data shown in the “learning” column in the middle, and the voice model is generated in the “test” column in the lower. This corresponds to a case where a speaker recognition test is performed using voice data uttered at the indicated time. Here, in the “learning” section in the middle, the utterance time of the voice data to be used and the total number of sentences are also shown. For example, in case B, the timings T1 to T3 (that is, the timing T
A voice model is generated using the voice data included in the text data for generating a total of 15 voice models uttered at 1, T2 and T3), and the voice data included in the test text data uttered at time T4 Is used to test speaker recognition.

Incidentally, the likelihood of the HMM is calculated by a likelihood normalization method based on the posterior probability (for example, the document "T. Matsuiand S. Furui," Likel
ihood normalization using a phoneme- and speaker-i
ndependent model for speaker verification ", Speech
Communication, Vol. 17, No. 1-2, pp. 109-116, 1995. "). Here, the threshold value is set so that the impersonator error rate calculated from the test data is equal to the false rejection rate, and the evaluation is performed based on the equal error rate.

When an experiment is performed under such conditions, FIG.
The result shown in FIG. This figure shows the average speaker verification error rate of each case and the average of all cases.
In the figure, “this method” is the method described in the above embodiment,
The "conventional method" is a method of maximum likelihood estimation of the speaker's HMM using all data available in each case, and the "one time" is using only the latest data (speech data for 5 sentences) in each case. This is a method of estimating the maximum likelihood of the speaker's HMM, and the corresponding column in each case shows the result of each of these methods. From these results, it can be seen that the present method is more effective than the conventional method, and that the performance is better when using voice data uttered at a plurality of times. In particular, looking at the average error rate of all cases, the average error reduction rate by this method is 12.5% for the conventional method and 70.8% for the one-time method. The effectiveness of this method is remarkably demonstrated.

The variance of the HMM generated by the present method is shown in the above equation (6), which is smaller than the variance of the HMM generated by the conventional method. HMM variance vector σ _jk ^(t) generated by the conventional method
Is the observation vector o ^(t) (n) with state j and Gaussian distribution k
Is defined as γ ′ ^(t) _jk (n), and μ _jk ^(t)平均 as an average vector adapted to the data at time t.

(Equation 7) (However, Equation 7 represents each component.) Based on this, the variance in each case is obtained, and
FIG. 5 shows a comparison with the variance of the HMM generated by this method.

FIG. 5 shows the average of the ratio of each element of the variance vector of each of the conventional method and the present method, the temporary method and the present method. Note that the variance vector in the one-time method is
In the above equation (7), it can be obtained by considering that the HMM is generated from only the latest data as T = 1 or the like. As shown in the figure, when the present method is compared with the conventional method, the conventional method has a larger variance, and the tendency increases as the number of data periods used increases. On the other hand, when this method is compared with the one-time method, the ratio of the variance values does not depend on the number of data periods used. This means that according to the present method, the spread of the variance of the HMM can be appropriately suppressed.

Equations (6) and (7) can be interpreted as averaging the variance value at each time (probabilistically) with a weight corresponding to each data length. However, this method is based on the idea that the data of voices uttered by a speaker at multiple periods are obtained from a plurality of populations corresponding to each period. For variance, Equation 6
Is closer to the uniform minimum variance unbiased estimator than that of equation (7). Due to this, as described above, the spread of HMM dispersion can be suppressed by the present method.

[0052]

As described above, according to the present invention, a parameter extracted from a component independent of time in a speaker's voice is estimated from a characteristic parameter of each time, and a voice is extracted based on the parameter. Since the model is generated, it is possible to generate a speech model based on a parameter having a small variance at each time, from which time-dependent fluctuation components have been removed. Accordingly, it is possible to generate a speech model that can maintain the discrimination ability and the matching ability at a high level with respect to the speaker's voice that fluctuates depending on the utterance time. Here, claim 2
In the invention described in the above, such a parameter having a small variance is estimated together with a conversion factor that converts the parameter with a variation corresponding to each time, and in the invention described in the third aspect, the parameter is estimated by converting the parameter with the conversion factor. It is estimated by the likelihood of the feature parameter at each time for the voice.

According to the invention described in claim 4 or the like in which speaker recognition is performed using the voice model, an effect is obtained that high recognition performance can be maintained over a plurality of periods.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a configuration of a speaker recognition device according to an embodiment of the present invention.

FIG. 2 is a diagram showing a procedure of generating a speech model in a model generation unit 3 of FIG.

FIG. 3 is a diagram showing utterance times of voice data used in learning and testing under the conditions of the embodiment.

FIG. 4 is a diagram illustrating the results of an experiment performed under the conditions of the example, using the average speaker verification error rate of each case and the average of all cases.

FIG. 5 is a diagram showing a comparison result of variances of HMMs generated by the present method, the conventional method, and the one-time method.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Feature parameter extraction part 2 Feature parameter storage part 3 Model generation part 4 Model storage part 5 Similarity calculation part 6 Speaker recognition judgment part 7 Threshold storage part

Claims (9)

[Claims] 1. A speech model generation method for extracting a feature parameter representing a speaker's personality from a speech and generating a speech model of the speaker based on the feature parameter, comprising the steps of: Each feature parameter is extracted, a parameter extracted from a component independent of the time in the voice is estimated from the feature parameter at each time, and a voice model of a speaker who utters the voice is generated based on the parameter. A method of generating a speech model. 2. A method for generating a speech model for extracting a feature parameter representing a personality of a speaker from a speech and generating a speech model of the speaker based on the feature parameter, comprising the steps of: A characteristic parameter is extracted, and a characteristic parameter of each time is extracted. From the characteristic parameter of each time, a parameter common to each time extracted from a component that does not depend on the time in the voice, and a conversion factor for converting the parameter with a variation corresponding to each time. And generating a voice model of the speaker who uttered the voice based on the parameters common to each time. 3. The method of generating a speech model according to claim 2, wherein for each pair of the estimated parameter common to each time period and the conversion factor, each parameter common to each time period is converted by a conversion factor and estimated. Generating a voice model based on a parameter common to each time period included in the group in which the likelihood is maximized; Method. 4. A speech model is generated in advance for each speaker using the speech model generation method according to any one of claims 1 to 3, and the speech model is generated from a speech uttered by an unspecified speaker. A speaker recognition method comprising: extracting a feature parameter; obtaining a similarity between the feature parameter and the speech model; and recognizing the unspecified speaker based on the similarity. 5. A feature parameter extracting means for extracting a feature parameter representing a speaker's personality from an input voice, and a time-independent component of the input voice from a feature parameter extracted from the input voice of a plurality of times. Model generation means for estimating a parameter extracted from the above, and generating a voice model of a speaker who uttered the input voice based on the parameter; model storage means for storing the generated voice model; Recognizing means for determining a similarity between a feature parameter extracted from an input voice uttered by a speaker and a stored voice model, and recognizing the unspecified speaker based on the similarity. Speaker recognition device. 6. A feature parameter extracting means, a model storage means, and a recognizing means according to claim 5, and a common parameter for each time extracted from a time-independent component of the input voice from the feature parameter for each time. And a model generating means for estimating a conversion factor for converting the parameter with a variation corresponding to each time, and generating a voice model of a speaker who uttered the input voice based on the parameter common to each time. A speaker recognition device comprising: 7. The speaker recognition apparatus according to claim 6, wherein the model generating means converts a parameter common to each time with a conversion factor for each set of the estimated parameter common to each time and a conversion factor. Determining the likelihood of the characteristic parameter of each period estimated for the input voice, and generating the voice model based on parameters common to each period included in the set having the maximum likelihood. Speaker recognition device. 8. A process of extracting characteristic parameters each representing a speaker's personality from input voices uttered at a plurality of periods, and from the extracted characteristic parameters of each period, independent of a time in the input voice. A recording medium storing a program for causing a predetermined arithmetic device to execute a process of estimating a parameter extracted from a component and a process of generating a voice model of a speaker who uttered the voice based on the parameter. 9. A process of extracting a feature parameter representing a speaker's personality from input speech, and extracting a feature-independent component of the input speech from a feature parameter extracted from the input speech of a plurality of periods. Estimating the parameters, and generating a speech model of the speaker who uttered the input speech based on the parameters; accumulating the generated speech model; and input speech uttered by the unspecified speaker. Recording a program for determining a similarity between the extracted feature parameter and the stored speech model, and causing a predetermined arithmetic device to execute a process of recognizing the unspecified speaker based on the similarity; Medium.