JP3092788B2 - Speaker recognition threshold setting method and speaker recognition apparatus using the method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention is used, for example, to identify the same person as a personal identification number based on an input voice, and to convert the input voice into an expression form using feature parameters. In the speaker recognition method for recognizing a speaker who uttered the input voice, a similarity between the input voice in the expression and the model of the voice in the above-mentioned expression registered in advance for the speaker is obtained. The present invention relates to a method for setting a threshold value used for speaker determination and a speaker recognition device to which the method is applied.

[0002]

2. Description of the Related Art FIG. 3 shows a functional configuration of a conventional text independent speaker recognition apparatus. First, a speaker is registered. A voice (registration voice) such as a sentence uttered for each speaker is input from the input terminal 11 to the feature parameter extracting unit 12.
Is converted into an expression format using feature parameters (for example, cepstrum, pitch, etc.) included in the voice, and the model creation unit 13 converts the time-series registration voice data of the feature parameter Models, such as Hidden Markov Model:
Notated as HMM. For example, a plurality of Gaussian distributions are represented by weighted addition). As a method of creating an HMM, for example, a document “Tomoko Matsui, Sadahiro Furui:“ VQ, Comparative Study of Text-Independent Speaker Recognition Method Using Discrete / Continuous HMM ”, IEICE Speech Research Group, SP91-8
9, 1991 "can be used. The HMM for each speaker obtained in this manner is registered in the model storage unit 14 for each speaker.

When recognizing a speaker, the uttered voice of the speaker is input from an input terminal 11 to a characteristic parameter extracting means 12, and is converted into a time series of characteristic parameters. Is similarity calculating means 15
Calculates the similarity between each speaker stored in the model storage unit 14 and the HMM.
In step S6, the input voice is compared with a threshold value stored in the threshold value storage unit 17 in consideration of a range of similarity fluctuation that can be regarded as a person's voice. The voice is determined to be the voice of the registered speaker of the HMM used for the calculation, and if smaller than the threshold value, it is determined to be voice of another person, and this determination result is output.

Conventionally, to set a threshold value,
Two error rates were considered, the rejection rate and the false acceptance rate. The rejection rate is obtained from the results of a speaker recognition experiment using the registration voice of the individual out of all the registered voices, and indicates the rate of false rejection of the individual. It is obtained from the result of a speaker recognition experiment using speech, and indicates the rate at which impostors are incorrectly accepted.
Depending on the purpose of speaker recognition, the false rejection rate may be more important than the false acceptance rate, and vice versa.
When the purpose was not clear, Bayes' theorem was used to determine the optimal threshold value (equal error rate threshold) at which the false rejection rate and the impostor acceptance rate were equal and gave the equal error rate.
As shown in FIG. 4A, the curve 21 indicating the rejection rate increases as the threshold value increases. On the other hand, the curve 22 indicating the impostor acceptance rate decreases as the threshold value increases. Conventionally, the impostor is a registered speaker other than himself / herself, and the similarity to each model (HMM) is calculated and speaker recognition is performed using all registration voices. The threshold was changed, that is, a speaker recognition experiment was performed, and the rejection rate curve 21 and the impostor acceptance rate curve 2 shown in FIG.
Intersection of the 2, that is set as a threshold seeking threshold phi ₀ both error rate becomes equal epsilon _0, ie sets the threshold for equal error rate by a registered voice.

[0005]

However, the model of the person himself
Is sufficiently robust against utterance differences, utterance fluctuations, etc.
If not, use your own model and create it
The similarity to the voice of the person (voice for registration of the person)
Compared to the similarity between the model and the voice spoken during recognition
All are generally large. Therefore, the person himself
The person who changes the threshold for the voice uttered when
When the rejection rate curve is obtained, for example, the dotted line curve 23 in FIG.
Thus, the rejection rate curve 21 using the registration voice of the person
Rejection rate is worse than
Rejection rate increases. In other words, the equal error rate due to the registration voice
Threshold value φ₀Is the threshold of equal error rate by speech for recognition.
Value φ ₀Value, and as a result,
Equal error rate threshold φ₀If used for recognition
There was a problem that the rejection rate increased.

In addition, since the amount of voice data for registration of a person is not so large, the model of the person is often not sufficiently robust against differences in utterance contents, fluctuations in utterance, and the like. The lack of high demand was also a problem. Furthermore, the voice of the speaker fluctuates with each utterance, and in particular, a few
It fluctuates greatly in units of months. From this point, in order to maintain high recognition performance, it is desired that each speaker be uttered at regular intervals to update the model.
When the model is updated in this way, the personal rejection rate characteristics and the impostor acceptance rate characteristics also change. Therefore, it is desirable to reset the threshold value when the model is updated.

[0007]

According to the method of the first aspect of the present invention, the equal error rate when a speaker recognition experiment is performed using a registration voice with an impostor as a registered speaker other than the registered person. A value obtained by subtracting a predetermined value from the given threshold value is set as the threshold value, that is, a value that gives a false acceptance rate higher than the equal error rate threshold value is set. This higher impostor acceptance rate
It is higher than the impostor acceptance rate at the equal error rate threshold by about the upper limit of the system error rate of this speaker recognition method. With this configuration, the rejection rate does not become too large even if the model is not robust.

In the method according to the second aspect of the present invention, the model is periodically updated, and each time the model is updated, the impersonator is registered as a registered speaker other than himself using the update voice and the updated model.
A value obtained by subtracting a value smaller than the predetermined value and smaller according to the number of updates from the threshold of the equal error rate at the time of performing the speaker recognition experiment is set as a new threshold. With this configuration, as the model is updated, it becomes more and more robust against differences in utterance content, utterance fluctuation, etc., and the ideal equal error rate for the recognition speech when the ideal model is used. The threshold value is asymptotic from the value giving the higher impostor acceptance rate.

An example of updating the threshold every time the model is updated will be described below. That is, the threshold value φ is set according to the following equation. φ = wφ ₁ + (1−w) φ ₀ (1) where φ ₀ is an equal error rate when a speaker recognition experiment is performed using a registration voice with the impostor as a registered speaker other than himself / herself. threshold, i.e. represents the initially set threshold, phi ₁ is based on the relationship (FIG. 4A) of the false acceptance rate and the threshold, false acceptance rate {equal error rate ε ₀ + x}% (Eg x
= 1%). This ｛equal error rate ε
₀ + x｝% corresponds to the upper limit of the impostor acceptance rate (the system estimation error rate) estimated from the performance of the speaker recognition method. w is a parameter that controls the speed at which the threshold value gradually approaches the threshold value of the equal error rate in accordance with the updating of the speaker model, and can be defined by the following equation, for example.

W = 2 / (1 + exp (0.25t)) (2) where t is the number of updates of the speaker model (t = 0, 1,
2,...), And this equation is an equation obtained by experiment.
According to the formulas (1) and (2), at t = 0, the time when a speaker recognition device is made or when all the speakers to be recognized are new, that is, the registration voice , And is set to a threshold smaller by Δφ (= φ ₀ −φ ₁ ) than the equal error rate threshold φ ₀ . Usually, as the number of updates of the model increases, w decreases, and Δφ also decreases and approaches φ ₀ . Note that φ
_{0, φ 1} is also determined for each model update.

[0011]

FIG. 1 shows the order of processing in an embodiment of the method of the present invention, and FIG. 2 shows the functional configuration of an embodiment of the apparatus of the present invention by assigning the same reference numerals to parts corresponding to those in FIG. . 2, in this embodiment, a feature parameter time storage unit 25 for temporarily storing a time series of feature parameters of each input voice at the time of registration and model update, and a model update in the model storage unit 14 when there is a model update instruction. And a threshold calculating means 27 for calculating a threshold at the time of registration and updating of the model and updating the threshold of the threshold accumulating unit 17.

When a voice for registration or a voice for update is inputted to the input terminal 11, it is converted into a time series of characteristic parameters by the characteristic parameter extracting means 12 as shown in FIGS. 1 and 2 (S ₁ ). model is the voice models in creation means 13 creates, at the time of model update, the feature parameter time series of update audio, updating the corresponding models in the model storage unit 14 is carried out (S _2).

In order to update the model, the time series of each characteristic parameter of the speech for registration and the speech for update for each speaker are held, and the time series of all the held speeches and the newly input update time are updated. A new model is created using the time series of the voice for use, and the corresponding model in the model storage unit 14 is updated. Alternatively, when the model is an HMM, the HMM of the speaker corresponding to the time series X of the feature parameters of the update speech is obtained by Bayesian estimation.
Estimate the parameter vector θ of the HMM that maximizes the product of the likelihood f (X | 1θ) with respect to the prior probability density function g (θ) reflecting the features of the voices uttered up to then.
Let θ be a new HMM.

Next, at the time of registration, voice for registration, at the time of model update
Using the update speech ₀And its threshold
φ₀Is calculated (S_Three). In other words, the feature parameters of these voices
The time series of the meter is temporarily stored in the feature parameter time storage unit 25.
Are stored in the storage unit 14 and the model of each model in the model storage unit 14.
The similarity is calculated by the similarity calculating means 15, respectively.
For the similarity, various thresholds are determined by the speaker recognition determining means 16.
Judgment is made for the value, and the impostor is regarded as a registered speaker other than himself.
Speaker recognition using registration voice (or update voice)
Test and the rejection rate curve shown in FIG.
And the error rate ε where both error rates are equal₀And its
Time threshold φ₀Ask for.

Thereafter, a threshold φ _{1 at} which the impostor acceptance rate gives (ε ₀ + x)% is obtained (S ₄ ), and a new threshold φ
Is calculated by calculating wφ ₁ + (1−w) φ ₀ (S ₅ ).
This new threshold value φ is set as the threshold value of the corresponding speaker in the threshold value storage unit 17. Next, the number of model updates t is incremented by 1 to end the process (S ₆ ). Steps S ₃ , S ₄ , S ₅ , S
Step ₆ is performed by the threshold value calculating means 27.

In FIG. 1, when t = 0, a threshold value is calculated using the voice for registration at the time of registration, and the impostor who is x% higher than the error rate ε ₀ which gives the threshold of the equal error rate at that time is calculated. threshold phi ₁ as the acceptance rate is set as a threshold, model may not robust, does not false rejection rate is too high. In addition, each time a model update is performed, a speaker recognition experiment is performed on the updated model using the updated voice and the impostor as a registered speaker other than the person itself. , Which is close to the equal error rate threshold for the model approaching, and w is small, and the difference from the ideally close equal error rate threshold is small. The more the value is set, that is, the more the model update is repeated, the more desirable the threshold becomes.

[0017]

Next, an experimental example for confirming the effect of the present invention will be described. The experiment was conducted by 20 men for about 15 months5
Sentence data uttered at two times (time A, B, C, D, and E) (one sentence length is 4 seconds on average) is targeted. Ten males were used as registered speakers and the other ten males were used as impostors. These voices are used as feature values,
That is, the cepstrum is converted into a time series for each short time.
The cepstrum has a sampling frequency of 12 kHz and a frame length of 3
2 ms, frame period 8 ms, LPC analysis (Linear Pre
(dictive coding, linear prediction analysis).
Ten sentences uttered at time A were used for registration. For the update, ten sentences uttered at time B were used as the first update, and ten sentences uttered at time C were used as the second update. In the test, five sentences uttered at times D and E were used one sentence at a time, that is, each model at times A, B, and C performed a test on the threshold value five times each. In the setting of the threshold value, x = 1%.

The effect of the present invention is that the text-independent type (for example, the documents "Tomoko Matsui, Sadahiro Furui:" VQ, discrete / continuous HM
M-Comparison of Text Independent Speaker Recognition Methods ", IEICE Spoken Language Study Group, SP91-89, 199
1)). The HMM for each speaker is
Weighted addition of 64 Gaussian distributions in one state (for example, literature "Tomoko Matsui, Sadahiro Furui:" Comparison of VQ, text-independent speaker recognition method using discrete / continuous HMM ", IEICE Speech Research Group) Documents, SP91-89, 199
1 ").

The results were evaluated by the average of the rejection rate and the false acceptance rate. The result is shown in FIG. 4B. In the conventional method, a result obtained by performing a speaker recognition experiment using all registration voices with the impostor as a registered speaker other than the registered person is represented by a threshold value of an equal error rate. From this, the method of the present invention, compared with the conventional method,
It turns out that it shows high performance. From these results, the method of the present invention was proved to be effective.

[Brief description of the drawings]

FIG. 1 is a flowchart showing a processing procedure in an embodiment of the method of the present invention.

FIG. 2 is a block diagram showing a functional configuration of an embodiment of the apparatus of the present invention.

FIG. 3 is a block diagram showing a functional configuration of a conventional speaker recognition device.

FIG. 4A is a diagram showing a relationship between a rejection rate and an impostor acceptance rate with respect to a threshold, and FIG. 4B is a view showing an experimental result for explaining an effect of the present invention.

Continuation of the front page (56) References JP-A-8-123475 (JP, A) JP-A-62-209500 (JP, A) JP-A-47-41103 (JP, A) JP-A-7-248791 (JP) , A) JP-B-63-29279 (JP, B2) IEEE Transactions on Acoustics, Speech and Signal Processing, Vol. ASSP-29, no. 2, April 1981, S.M. Furui, “Cepstral Analysis Technique for Automatic Speaker Verification”, p. 254-272 Proceedings of the Spring Meeting of the Acoustical Society of Japan 1996, I-I, 1-5-6, Tomoko Matsui Examination of updating method of model and threshold ”, p. 11-12, (issued March 26, 1996) IEICE Technical Report [Voice], Vol. 94, no. 90, SP94-22, Tomiko Matsui et al. “Normalization of speaker matching likelihood by phoneme / speaker independent model”, p. 61-66 (issued on June 16, 1994) IEICE Technical Report [Voice], Vol. 95, No. 468, SP95-120, Tomoko Matsui et al., “Method of updating model and threshold in speaker verification,” p. 21-26 (issued on January 19, 1996) Transactions of the Institute of Electronics, Information and Communication Engineers, Vol. J 81-D- ▲ II ▼ No. 2, Febuary 1998, Tomoko Matsui et al., "Method of updating models and thresholds in speaker verification," p. 268-276 (Issued February 25, 1998) (58) Field surveyed (Int. Cl. ⁷ , DB name) G10L 15/00-17/00 JICST file (JOIS)

Claims (4)

(57) [Claims] 1. An input speech is converted into an expression form using feature parameters, and a similarity between the input speech in the expression form and a speech model in the above-mentioned expression form registered in advance for a speaker is obtained. A method for setting the threshold value in a speaker recognition method for recognizing a speaker who has uttered the input voice by comparing the similarity with a speaker determination threshold value, comprising the steps of: Calculate the two error rates, the rejection rate and the impostor acceptance rate, using the voice uttered at this time and the above registration model, and subtract a predetermined value from a threshold value at which the calculated two error rates become equal. A threshold value for speaker recognition, wherein the threshold value for speaker determination is set to the threshold value. 2. Each time the model for each speaker is updated, the two error rates are calculated using the updated model and the voice uttered at the time of the update, and the calculated two error rates are calculated. 2. The speaker according to claim 1, wherein the threshold value for speaker determination is updated to a value obtained by subtracting a value smaller than the predetermined value and smaller than a previous value from a threshold value at which an error rate becomes equal. How to set the threshold for recognition. 3. The method according to claim 1, wherein the predetermined value is substantially equal to an upper limit of an error rate of the speaker recognition method itself. 4. An input speech is converted into an expression format using feature parameters by a feature parameter extraction unit, and a model of the input speech in this expression format is created by a model creation unit and stored in a model storage unit. The similarity of each of the models in the model storage means is calculated by the similarity calculation means for the speech of the expression format converted by the feature parameter extraction means, and these calculated similarities are stored in a threshold storage section. The threshold value indicating the range of variation of the similarity that can be regarded as the voice of the individual is compared with the speaker recognition determining means. If the similarity is higher, the voice is the voice of the individual, and if the similarity is lower, the voice is the voice of another person. In the speaker recognition device, when a model update instruction is issued, the model in the model storage unit of the corresponding speaker is updated by an input speech in an expression form using the feature parameter from the feature parameter extracting unit. Means for updating the model, and calculating a rejection rate and an imposter acceptance rate for the speech at the time of updating the updated model, and determining whether the rejection rate and the acceptance rate are equal. Threshold calculating means for updating a threshold value of a corresponding speaker in the threshold accumulation unit to a value obtained by subtracting a slightly smaller value from the threshold value. Accompanying speaker recognition device.