JP3448371B2 - HMM learning device - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a Hidden Markov Model (hereinafter, referred to as a probabilistic model) which is a probabilistic model that approximately expresses statistical characteristics of speech by a distribution such as Gaussian distribution.
It is called HMM. ) Learning device.

[0002] [Detailed Description of the Invention]

[0003]

2. Description of the Related Art In recent years, a voice recognition device using an HMM has been actively developed. This HMM is a model of statistical characteristics of speech obtained from a large amount of speech data. This model can (1) statistically process fluctuations of utterance in the form of distribution, (2) depending on the speaker It has the advantage of being able to absorb differences in utterance duration.

A case will be described as an example where voice recognition of a word is performed using a phoneme HMM having these advantages.

Generally, when an HMM is created in units of phonemes so that a word is formed by a unit smaller than that, for example, phonemes are connected, the word recognition for an arbitrary word is performed by connecting the phoneme HMMs. Can be done. FIG. 4 is a conceptual diagram for performing word recognition based on the phoneme HMM.

[0006] Now, the recognition targets registered in the dictionary are "Ukesu (U / CH / I / K / E / S / U)" and "Uchidake (U / CH / I / A / W / A / S / E) ”and“ Uru (U / R / U) ”are three words, the phoneme HMM that needs to be created is“ U / CH / I / ”that appears in the dictionary.
It is understood that only 9 types of "K / E / S / A / W / R" are sufficient.

Therefore, in word recognition, the phoneme H
The word HMM corresponding to the word existing in the dictionary is created by connecting the MMs, and the one close to the input speech (word) can be obtained as the stochastic likelihood (probability). .

As described above, by learning the voice information of a large number of speakers and creating the phoneme HMM in advance, it is possible to recognize even when the input voice is a word,
The above is the outline of the HMM.

By the way, such an HMM is generally created by using several hundred words for learning. However, uttering hundreds of words by the user is not realistic considering the burden on the user. In order to avoid such a point, there is a speaker adaptation method as a method of tuning the HMM to a user's voice feature by using a small number of learning words, and this speaker adaptation method is based on the Institute of Electronics, Information and Communication Engineers D-2 Vol. J76
-D-2 No. 12 December 1993, pp. 2469-2476.

A speech recognition method using the HMM will be described below with reference to FIGS. 5 to 9.

FIG. 5 is a voice pattern showing the relationship between the logarithmic power (hereinafter simply referred to as power) of the input voices "ba (ba)" and "bu (bu)" and time. This voice pattern is obtained by dividing the voice band of the input voice by, for example, 16 band filters, performing frequency analysis of the voice, and then taking logarithmic power for each time. As can be seen from FIG. 5 (b), there is a difference in power change even in the same phoneme / b / section,
Looking at the power change in the section b /, the power change in the first part of / b / is small and gradually increasing. Focusing on the power change, the phoneme / b / There are few changes, but there are many fluctuations: section 2; power rising portion; section 3; steep power rising portion;

On the other hand, FIG. 6A shows the phonemes of FIG.
It is the figure which divided the section of b / into the section 1 to the section 3. Further, FIG. 6B corresponds to the phoneme / b / sections 1 to 3 in FIG. 6A, respectively, and shows a histogram of the power change amount, which is approximated by a Gaussian distribution. Generally, in the HMM, the characteristics of the voice are expressed as such distributions 1 to 3. For example, when a voice is analyzed by a 16-channel bandpass filter or the like, one Gaussian distribution is obtained for each channel. Here, by considering these 16 Gaussian distributions as one component, the average value of each of the 16 Gaussian distributions included in this component can be expressed as a vector, and such a vector will be referred to as an average vector hereinafter. .

By the way, FIG. 7 is a schematic configuration diagram of a conventional HMM learning device based on speaker adaptation of the HMM, and a speech recognition device using this learning device.

In FIG. 7, reference numeral 1 is a voice analysis unit for analyzing the characteristics of the input voice for each frequency band, and 2 is an initial model storage unit for storing an initial model of the HMM by learning, and the initial model is specified. It may be an HMM of a specific speaker created by using the voices of the speakers, or an HMM of an unspecified speaker learned by using a large number of speaker voices. As a specific learning method, a well-known forward backward algorithm or a learning rule based on Viterbi alignment may be used.

A learning unit 3 re-learns the above-mentioned initial model by using an input voice.
It is assumed that only the average vector is learned from among the parameters expressing.

The learning of the average vector in the learning section 3 will be described below.

Of the components in the initial model of the HMM
The set of average vectors is C^R(C^R= (C ₁ ^R,,, c_k ^R・ ・ ・
., C_m ^R), Where m is the number of all components
Represent ), This average vector set C^RIs learning
The analysis result analyzed by the voice analysis unit 1 in the unit 3
Re-learned using the average vector C after re-learning^IGot
Be done. Where C^I= (C₁ ^I,,, c_k ^I..., c_m ^I )so
is there. Where c_k ^RAnd c _k ^IShall correspond.
That is, c_k ^RIs c after learning_k ^Ibecome.

Reference numeral 4 denotes a speaker adaptation unit for making the HMM re-learned in the learning unit 3 into a more accurate model. This speaker adaptation unit 4 calculates the movement vector and relates to a phoneme HMM that has not been learned. This is a part that performs interpolation / interpolation processing on the moving vector v _n related to the average vector c _n ^R , and further performs smoothing processing on all the moving vectors to adapt the average vector.

Reference numeral 5 denotes a post-adaptation model storage unit for storing the HMM after the adaptation by the speaker adaptation unit 4.

The above is the configuration of the conventional HMM learning apparatus, and the processing in the speaker adaptation unit 4 will be described here. (A) Calculation of movement vector v _k The difference between the initial model C ^R and the average vector of each component in C ^I after re-learning is calculated according to the following equation. Hereinafter, this is called a movement vector.

V _k = c _k ^I −c _k ^R (where k = 1, 2, ..., ^M ) (B) Interpolation / interpolation processing of the movement vector v _n FIG. The mean vector of the initial model (c ₁ ^R ,
c ₂ ^R , c ₃ ^R ) is re-learned, and after this re-learning, it becomes a mean vector (c ₁ ^I , c ₂ ^I , c ₃ ^I ). Further, the average vector c _n ^R represents that learning has not been performed because there is no phoneme related to c _n ^R in the learned speech. (V ₁ , v ₂ , v ₃ ) in the figure are movement vectors obtained from the learned components.

Here, the movement vector interpolation / interpolation processing is to obtain the movement vector v _n related to the unlearned average vector c _n ^R as shown in FIG. 8B.

FIG. 8B is a conceptual diagram for calculating the movement vector v _n . This movement vector v _n is calculated by interpolating based on the movement vectors v ₁ , v ₂ and v _3. And the movement vector v _n is the movement vector v ₁ , v
It can be represented by a weighted average of ₂ and v ₃ . (C) Movement Vector Smoothing Process FIG. 9 is a conceptual diagram of (B) movement vector smoothing process performed after the interpolation / interpolation process of the movement vector v _n .

As described above, the estimated and calculated movement vector is considered to include a large amount of estimation error when the learning is not performed with a sufficient number of words. Therefore, it is considered that the direction of the movement vector calculated from the one including such an estimation error has a discontinuous movement. In FIG. 9, the discontinuous movements here mean that the movement vectors v ₁ , v ₂ , and v ₃ are all in substantially the same direction, but v _k is a direction different from those three movement vectors. That is facing.

Therefore, the moving vectors v ₁ , v ₂ , and v with respect to the mean vector c _k ^R and the mean vectors in the vicinity thereof are
Based on ₃ , the smoothed motion vector v _k ^S is calculated by making a correction. That is, the movement vector v _k
Is slightly corrected to the left under the influence of the movement vectors v ₁ , v ₂ and v ₃ . This correction processing is called smoothing processing. (D) Average vector adaptation processing Using the smoothed movement vector v _k ^S and the average vector C _k ^R obtained in the above (C), the average vector C _k ^S of the HMM after speaker adaptation according to the following equation. Calculate (k = 1, ..., m).

C _k ^S = C _k ^R + v _k ^S

[0027]

However, as described above, in the conventional HMM creation, (B) interpolation processing by interpolation of the moving vector and (C) smoothing processing of the moving vector are performed. However, when the learning audio material, that is, the word speech is small in number, the average vector learned by the learning unit 3 is small, and the movement vector obtained from the learned small average vector is also small. In such a case, that is, movement vectors v ₁ , v ₂ , and v ₃ ,.
.. is small, the number of vectors to be interpolated using the movement vectors v ₁ , v ₂ , and v ₃ , ... is large, and in the smoothing process, it is obtained by learning. Since the smoothing process is performed from a small number of vectors, there is a problem that only an inappropriate model can be obtained as a model of the input speaker.

[0028]

Therefore, the present invention has been made in view of the above problems, and a small amount of the speaker subspace movement vectors v _i , _s , _m ^{n of} a plurality of representative speakers can be obtained. input speaker of the speaker subspace movement vector v _i of the input speaker obtained from the learning audio material, _{s, m} ^inp to the distance to close representative speaker of the speaker subspace movement vector v _{i, s, m} select ^SPNO, characterized in that to adapt the surrogate table speaker speaker subspace movement vector v _{i, s,} to the input talker the unspecified speaker HMM by modifying the _m ^SPNO.

Further, according to the present invention, a voice analysis unit (1) for analyzing characteristics of an input voice, an initial model storage unit (2) for storing an initial model of an HMM, and an input speaker in the voice analysis unit (1). Initial model storage unit using the results of analyzing the human voice (2)
A learning unit (3) for learning the HMM stored in
(3) mean vector mu _{i, s} of the HMM input speaker learned in, _m ^inp and initial model mean vector mu _{i, s} of the HMM stored in the storage unit (2) is determined from the difference between _m input speaker of the speaker subspace movement vector v _i is calculated using the difference vector, _s, speaker subspace movement vector calculating unit input speaker calculating the _m ^inp and (10a), of the input speaker speaker subspace moving vector calculation unit speaker subspace of the input speaker obtained by (10a) movement vector v _{i, s,} speaker subspace motion vector storage unit of the input speaker for storing _m ^inp (10
b) and the speaker subspace movement vector v _i , _s , _m ⁿ of the representative speaker
Speaker subspace movement vector memory of the representative speaker
(12) and said input speaker of the speaker subspace motion vector storage unit (10b) to store the input speaker of the speaker subspace movement vector v _{i, s, m} ^inp and distance to close representative speaker Speaker subspace movement vector v _i , _s , _m ^spno of the representative speaker, and the speaker subspace movement of the representative speaker obtained by the representative speaker selection unit (10 c). vector v _i, with _{s, m} ^SPNO, speaker subspace movement vector v _i of the input speaker, _{s, m} ^inp, and mean vector mu _{i, s} of the initial model, the _m, the average vector after the speaker adaptation mu _{i, s,} speaker adaptation after model construction unit for determining the _m ^inp and (10d), the mean vector mu _{i, s} after the speaker adaptation, _m
and a post-adaptation model storage unit (14) for storing ^inp .

[0030]

The average vector of the components in the initial model of the HMM is learned by using the analysis result of the input voice analyzed by the voice analysis unit (1).

After that, by using the difference between the average vector of the components in the initial model and the average vector after retraining corresponding to this average vector, the speaker subspace movement vectors v _i , _s , _m of the input speaker are used. Calculate ^inp .

The speaker part of the representative speaker, which is distance-wise close to the speaker subspace movement vector v _i , _s , _{m of} the input speaker stored in the speaker subspace movement vector storage unit (10b) of the input speaker. Select spatial movement vectors v _i , _s , _m ^spno .

Representative talk obtained by the representative talker selection unit (10c)
Speaker's subspace movement vector v_i， _s，_m ^spnoInput story
Speaker's subspace movement vector v_i，_s，_m ^inp, And early
Model mean vector μ_i，_s，_mAfter speaker adaptation using
Mean vector of_i，_s，_m ^inpAfter speaker adaptation model building department
Find in (10d).

[0034] Finally, the mean vector mu _{i, s} after the speaker adaptation, and stores the _m ^inp adaptive after the model storage unit (14).

[0035]

Embodiments of the present invention will be described with reference to FIGS. A mixed Gaussian distribution type of a diagonal covariance matrix is used for the HMM.

FIG. 1 is a schematic configuration diagram of an HMM learning device according to the present invention, and FIG. 2 is a detailed configuration diagram centering on a speaker adaptation unit 10 of the HMM learning device according to the present invention. The same components as those of the conventional HMM learning device are designated by the same reference numerals.

The first difference in the configuration of the learning device for the HMM of the present invention from the conventional one is that a speaker adaptation unit 10 is provided instead of the speaker adaptation unit 4, and this speaker adaptation unit 10 is The input speaker includes a speaker subspace movement vector calculation unit 10a, an input speaker speaker subspace movement vector storage unit 10b, a representative speaker selection unit 10c, and a post-adaptation model construction unit 10d.

The second difference in the configuration of the learning device of the HMM of the present invention from that of the conventional one is that the speaker subspace movement vector storage unit 12 of the representative speaker is connected to the representative speaker selection unit 10c.
And the speaker subspace movement vector calculation unit 11 of the representative speaker.

The third point in which the configuration of the learning device of the HMM of the present invention is different from the conventional one is that the initial model storage unit 2, the speaker subspace movement vector storage unit 12 of the representative speaker, and the speaker adaptation unit 10 are provided. That is, the post-adaptation model creation unit 13 that creates an HMM based on the HMM and the post-adaptation model storage unit 14 that stores the HMM are provided.

Here, (A) a speaker subspace movement vector calculation unit 11 of a representative speaker and (B) a speaker subspace movement vector storage unit 12 of a representative speaker, which are typical constituent features of the present invention. , (C) speaker subspace movement vector calculation unit 10a of the input speaker, (D) speaker subspace movement vector storage unit 10b of the input speaker, (E) representative speaker selection unit 10c, (F)
Functions of the speaker-adapted model construction unit 10d, the (G) -adapted model creation unit 13, and the (H) -adapted model storage unit 14 will be described in detail. (A) Speaker Subspace Movement Vector Calculation Unit 11 of Representative Speaker This calculation unit 11 has a function of obtaining speaker subspace movement vectors of a plurality of representative speakers. Here, the speaker subspace movement vector is obtained by using the difference between the initial model and the mean vector of the Gaussian distribution of the HMM after retraining the initial model, and the speaker part is calculated in the following steps. The spatial movement vector can be obtained.

Step 1: The initial model (λ) stored in the initial model storage unit is used as the initial model of the phoneme HMM of each representative speaker. Here, I represents the number of the phoneme HMM, and since 39 phoneme HMMs are used in this embodiment, I = 39. Λ _i is the i-th phoneme HM
M is shown.

Λ = {λ ₁ , ..., λ _i , ..., λ _I } In addition, λ _i is λ _i = {w _i , _s , _m , a _i , _s1 , _s2 , μ _i , _s ，
It is represented by _m , σ _i , _s , _m }.

Note that w _i , _s , _m , μ _i , _s , _m , and σ _i , _s ,
_m represents a weight, a mean vector, and a vector of variance values for the m-th Gaussian distribution in the s-th state of the i-th phoneme HMM, respectively. a _i , _s1 , and _s2 represent the transition probabilities from the s1th state to the s2th state of the i-th phoneme HMM. Since a 33-dimensional vector is used as the feature quantity in this embodiment, μ _i , _s , and _m , Σ _i , _s , and _m are 33-dimensional vectors.

Here, a specific speaker model or an unspecified speaker model may be used as the initial model.

By the way, μ _i , _s , and _m are C _k shown in the conventional example.
^It is the same as ^R, and the symbols μ _i , _s , and _m are used hereinafter for convenience of description in the present embodiment.

Step 2; The HMMs of the representative speaker are connected so as to correspond to the phoneme sequence of the input voice of the representative speaker, and learning is performed. Learning is performed only for w _i , _s , _m and μ _i , _s , _m , and λ _i ⁿ = {w _i ⁿ , _s , _m , as an ^nth speaker model.
a _i , _s1 , _s2 , μ _i ⁿ , _s , _m , σ _i , _s , _m } are obtained. Here, n represents the number of the representative speaker, n = 1, 2, ...
N, and in this embodiment, 30 representative speakers were used, so N = 30.

[0047] Step 3: For each representative speaker, the difference t _i ⁿ of the average value, _s, a _m ask.

∀ _i , _s , _m ∈Ωt _i , _s , _m ⁿ = μ _i , _s ,
_m ⁿ −μ _i , _s , _m (n = 1, 2, ..., N) where Ω is a set of subscripts _i , _s , and _m of the average vector μ _i , _s , _m included in λ. Represents

[0049] Here, the difference t _i ⁿ of the mean, _s, and _m is the same as the motion vector shown in the conventional example.

Step 4; According to equation 1, the speaker subspace movement vectors v _i , _s , _m ⁿ of the representative speaker are obtained. here,
It is assumed that a speaker subspace movement vector is obtained for each subspace by using K average vectors that are close in distance to μ _i , _s , and _m .

[0051]

[Equation 1]

Here, K _i , _s , _m is a set of subscripts relating to K average vectors in the vicinity of μ _i , _s , _m . Also,
D (a, b) represents the distance between the vectors a and b. f is a variable that controls the value of a fuzzy class function called fuzzyness. In addition to the fuzzy class functions, triangular windows, rectangular windows,
It is also possible to use a function such as Gaussian distribution.

[0053] On the other hand, t _i ^n, _s, may be used as the speaker subspace movement vector of the representative speaker _m.

Further, the learning is {w _i , _s , _m , a _i , _s1 , _s2 ,
Of μ _i , _s , _m , σ _i , _s , _m }, at least μ _i , _s , _m
You only have to learn to include. Of _{course, {w i, s, m} ,
All of a _i , _s1 , _s2 , μ _i , _s , _m , σ _i , _s , _m } may be learned. (B) Speaker subspace movement vector storage unit 12 of representative speaker The speaker subspace movement vector storage unit 12 of the representative speaker is a plurality of speaker subspace movement vector calculation units 11 of the representative speaker. Speaker subspace movement vector v of the representative speaker of
Memorize _i , _s , and _m ⁿ . (C) Speaker subspace movement vector calculator 10 of the input speaker
Based on the learned model by a learning unit 3, obtains speaker subspace movement vector v _i of the input speaker, _s, the _m ^inp the following steps. Here, inp represents the input speaker.

[0055] Step 1: calculate the difference t _i of the average value, _s, a _m ^inp.

[0056]

[Equation 2]

[0057] Step 2: As the number 3, the input speaker subspace movement vector v _{i, s,} a _m ^inp determined.

[0058]

[Equation 3]

Here, E represents a set of subscripts of the average vector of the phoneme HMM corresponding to the phonemes appearing in the learning speech material. (D) Speaker subspace movement vector storage unit 10 of the input speaker
b Input speaker speaker subspaces motion vector storage unit 10b stores the input speaker of the speaker subspace movement vector calculating unit input speaker subspaces motion vector calculated at ^{_{10a v i, s, m i}} np. (E) Representative speaker selection unit 10c Considering the branch probability of each component of the phoneme HMM,
According to Equation 4, the input speaker subspace movement vector v _i , _s , _m
^inp and distance to close representative speaker subspaces movement vector v _{i, s,} representatives speaker numbers with _m ⁿ (spno), and subspace representative speaker having a number (SPNO) of the representative speaker Select the movement vectors v _i , _s , _m ^spno .

[0060]

[Equation 4]

(F) Speaker-adapted model construction unit 10d In the speaker-adapted model construction unit 10d, the representative speaker selection unit 1 is used.
Speaker subspace movement vector representative speaker obtained in ^{_{0c v i, s, m spno}} , speaker subspace movement vector v _i of the input speaker, _{s, m} ^inp, and mean vector of the initial model mu _i , _S ,
with _m, the average vector mu _{i, s} after the speaker adaptation, seek _m ^inp.

[0062]

[Equation 5]

In this embodiment, W = 0.5 is set.
It was (G) Adaptation model creation unit 13 In the post-adaptation model creation unit 13, the speaker post-adaptation model construction unit
Mean vector μ after speaker adaptation constructed in 10d_i，_s，
_m ^inp, And the initial values stored in the initial model storage unit 2.
Weight w for the Gaussian distribution of the model_i，_s，_m, Transition probability
a_i，_s1，_s2And the variance value vector σ_i，_s，_mOr input speaker
Weight w for the Gaussian distribution of_i，_s，_m ^inp, Transition probability
a_i，_s1，_s2 ^inpAnd the variance value vector σ_i，_s，_m ^inpOr generation
Stored in the speaker space movement vector storage unit 12 of the speaker
Weight w for Gaussian distribution_i，_s， _m ^spno, Transition probability
a_i，_s1，_s2 ^spnoAnd the variance value vector σ_i，_s，_m ^spnoFor
And create a model after adaptation. (H) Adapted model storage unit 14 The post-adaptation model storage unit 14 is created by the post-adaptation model creation unit 13.
Store the created post-adaptation model.

Based on the flowchart shown in FIG. 3, learning processing and speaker adaptation processing for constructing a model after speaker adaptation in the speaker adaptation model construction unit 10d using the above-described configuration of the present invention are performed. Each step will be described below with reference to FIGS. 1 and 2.

The flowchart shown in FIG. 3 can be roughly divided into a learning process in the first stage (steps S1 to S5) and a speaker adaptation process in the second stage (steps S6 to S9).

First, in step S1, the phoneme HMM stored in the initial model storage unit 2 is used as the initial model of the input speaker. In step S2, the initial model storage unit 2
The word HMMs are created by concatenating the phoneme HMMs stored in. In step S3, the learning unit 3 learns the word HMM using the learning audio material. In step S4,
The learning unit 3 decomposes the word HMM into a phoneme HMM.
In step S5, by repeating learning, for example, it is determined whether or not a termination condition for converging the average vector of the phoneme HMMs in the learning speech material is satisfied, and if the termination condition is satisfied, the process proceeds to step S6. If the ending condition is not satisfied, the process returns to step S2.

[0067] In step S6, the speaker space motion vector calculation section 10a of the input speaker, obtains an average vector of the components of the initial model and the difference vector t _i between the mean vector of the phoneme HMM after learning, _s, the _m ^inp . In step S7, the speaker subspace movement vector calculating portion 10a of the input speaker, speaker subspace movement vector v _i of the input speaker, _s, the _m ^inp determined.

In step S8, the representative speaker selection unit 10c
At the speaker subspace movement vector v of the input speaker at_i，_s，
_m ^inpSubspace movement vector v of the representative speaker close to_i，_s，_m ⁿ
Select the number (spno) of the representative speaker with. Step
In step S9, in the speaker-adapted model building unit 10d,
Speaker portion of the representative speaker obtained by the representative speaker selection unit 10c
Space movement vector v_i，_s，_m ^spno, The speaker part of the input speaker
Space movement vector v_i， _s，_m ^inp, And the average of the initial model
Vector μ_i，_s，_m, The average vector after speaker adaptation
Le μ_i，_s，_m ^inpAsk for.

Here, learning was performed by the HMM learning device of the present invention, and the evaluation test was conducted.

As the initial model, an unspecified speaker model created from a part of audio data of 30 male speakers in the ASJ continuous speech database was used. 30 male speakers from the same database were used to create the representative speaker model. Evaluation was done by a male speaker 7 included in the Japanese common voice data of Jkyo.
100 words of 0 place names were used. Analysis conditions are sampling frequency 12 kHz, Hamming window length 21.3 ms, 16
Next LPC analysis, frame period 5 ms. A 16th-order LPC cepstrum, a 16th-order Δ cepstrum, and a 33-dimensional vector of Δ logarithmic power were used as the feature amount. HMM
Is a 4-state 3-loop, mixed Gaussian distribution type of diagonal covariance matrix, and the arc from each state is a tide arc. The number of models was 39.

A part of 100 place names (1 to 10 place names) was used as the speaker adaptation audio material. The place name other than the voice material used for speaker adaptation was used as the evaluation voice material. In addition, the evaluation was performed twice by changing the speaker adaptation word set.
The recognition word dictionary was 100 place names.

Table 1 shows word recognition results using the learning device of the present invention, conventional word recognition results that do not use the speaker adaptation method, and conventional word recognition results that use an unspecified speaker model (initial model). It is shown for each number of adaptive words.

[0073]

[Table 1]

As can be seen from this table, the recognition rate of the word recognition result using the learning device of the present invention is higher than that of the conventional word recognition result, indicating that the HMM learning device of the present invention is effective. I understand.

[0075] In the learning unit 3 of this embodiment, as in the conventional learning unit 3, the mean vector mu _{i, s} of the Gaussian distribution of the phoneme HMM unknown speaker, although relearn _m ^inp, However, the present invention is not limited to this, and the weights w _i , _s , _m regarding the Gaussian distribution, the vector of variance values σ _i , _s , _m , or any combination including the transition probabilities a _i , _s1 , _s2 may be retrained.

By the way, in step S8 shown in FIG.
Representative speaker selection section speaker subspace of the input speaker movement vector in 10c v _{i, s, m} ^inp and close representative speaker subspaces movement vector v _{i, s,} representatives speaker numbers with _m ⁿ (spn
Although the concept of selecting o) has been described, the present inventor proposes the following two methods as specific representative speaker selecting methods. That is, the first method is (1) a method based on the distance between speaker space movement vectors, and the second method is (2)
It is a method based on the likelihood of HMM for learning speech,
Each method is explained in detail below. (1) Method based on distance between speaker space movement vectors (a) Distance between speaker space movement vectors Methods using this speaker space movement vector include (a-
1) There is a case where all the speaker space movement vectors are used, and (a-2) a case where only the speaker space movement vector related to the learned phoneme HMM is used. Select the person.

(A-1) When all the speaker space movement vectors are used

[0078]

[Equation 6]

(A-2) When only the speaker space movement vector regarding the learned phoneme HMM is used

[0080]

[Equation 7]

Here, d _a is a distance measure expressed by equation (8).

[0082]

[Equation 8]

However, v _i , _s , _m , and _d ^inp represent the average value of the d-th element of the characteristic parameter in v _i , _s , and _m ^inp , and since a 33-dimensional vector is used as the characteristic amount in this embodiment, 1 ≦ d ≦ 33. W _d represents the weight. (B) Method Based on Distance Between Speaker Space Movement Vectors Considering Gaussian Distribution Variance Also in this method, (b-1) when all speaker space movement vectors are used and (b) -2) In some cases, only the speaker space movement vector related to the learned phoneme HMM is used, and the representative speaker is selected using the following formulas.

(B-1) When all the speaker space movement vectors are used

[0085]

[Equation 9]

(B-2) When only the speaker space movement vector related to the learned phoneme HMM is used

[0087]

[Equation 10]

Here, d _b is a distance measure expressed by the equation 11.

[0089]

[Equation 11]

Naturally, if the variance is re-learned when the initial model of each representative speaker and the initial model of the input speaker are re-learned, then Equations 9, 10 and 11 are respectively Equations 12, 13 and
And 14, respectively.

[0091]

[Equation 12]

[0092]

[Equation 13]

[0093]

[Equation 14]

However, σ _i , _s , _m , and _d ^inp represent the variance value regarding the d-th element of the characteristic parameter in σ _i , _s , and _m ^inp , and since a 33-dimensional vector is used as the characteristic amount in this embodiment, Similar to the above, 1 ≦ d ≦ 33.

(2) HMM-likelihood-based method for learning speech In this method, learning speech material (O) is converted into a phoneme HMM.
Likelihood L (Λ when recognition is performed using (λ ′)
_This is a method of selecting a representative speaker by _j (λ '), o _j ). Here, Λ _j (λ ′) represents a word model corresponding to the j-th learning voice o _j . The following two types were used as the phoneme HMM, and the likelihood was calculated for each. First of all, n
An HMM obtained by moving the mean vector μ _i , _s , _m ⁿ of λ ⁿ by the speaker space movement vectors v _i , _s , _m ⁿ of the th representative speaker
(Λ ⁿ ') using the likelihood ^{_{L (Λ j n (λ n}} '), o j) seek. Here, λ ⁿ 'i-th phoneme HMM of λ ^n' =
{W _i , _s , _m , a _i , _sj , _sk , μ _i , _s , _m ⁿ ',
σ _i , _s , _m ² }. Also, μ _i , _s , _m ⁿ '= v
_i , _s , _m ⁿ + μ _i , _s , _m , and the cumulative likelihood is calculated according to the following Expression 15.

[0096]

[Equation 15]

Next, based on the nth representative speaker,
Post-adaptation HMM (λ^{ad apt}(N))
Likelihood L (Λ_j ⁿ(Λ^adapt(N)), o_j). This
Where λ^adaptThe i-th phoneme HMM in (n) is λ_i ^adapt
(N) = {w_i，_s，_m, A_i，_{s j}，_sk, Μ_i，_s，
_m ^adapt(N), σ_i，_s，_m ²} Is represented. Also, μ
_i，_s，_m ^adapt(N) = Wv_i，_s，_m ^inp+ (1.0-W)
v_i，_s，_m ⁿ+ Μ_i，_s，_mAnd according to the following equation 16,
Cumulative likelihood is calculated.

[0098]

[Equation 16]

As described above, after the representative speaker is selected by the representative speaker selection method as described above, step S in FIG.
Processing similar to that of 9 is performed.

FIG. 10 shows the result of recognition using the representative speaker selected by the representative speaker selection method proposed this time.
Shown in.

As is apparent from FIG. 10, (1) In the method based on the distance between the speaker space movement vectors, the one considering the variance (the above-mentioned D _b1 and D _b2 correspond) is the speaker space movement. It can be seen that the recognition rate is improved as compared with the case of using the distance between the vectors (corresponding to D _a1 and D _a2 described above). Also,
Among them, it is better to use the distance between the speaker space movement vectors related to only the learned phonemes (corresponding to D _b2 described above).
Distance based on all speaker space movement vectors (D _b1 above
A slightly higher recognition performance is obtained compared to the case where (is applicable). It is considered that this is due to the interpolation error of the movement vector regarding the unlearned phoneme.

On the other hand, a method based on the distance between the speaker space movement vectors in consideration of the variance of the Gaussian distribution (the above-mentioned D _b1 , D
It can be seen that _b2 is applicable) and the performance is almost the same as the method based on the likelihood of HMM for learning speech (the above L ₁ and L ₂ are applicable). Further, the method based on the distance between the speaker space movement vectors has an advantage that the amount of calculation is smaller than the method based on the likelihood of the HMM for the learning voice.

From the above results, it was found that the recognition rate is better when the variance of the Gaussian distribution is considered in the method based on the distance between the speaker space movement vectors.

[0104] Furthermore, fuzzy grade function (e.g., F _i indicating the number _{1 ', s', m'} ) with even distance measure D is used considered better to use a distance measure d in consideration of dispersion becomes better recognition rate To be

[0105]

As is apparent from the above description, according to the present invention, the speaker subspace movement vectors v _i , _s , _m ^spon , and the input speaker of the representative speaker obtained by the representative speaker selection unit. speaker subspace movement vector v _{i, s, m} ^inp, and mean vector mu _{i, s} of the initial model, with _m, the average vector mu _{i, s} after the speaker adaptation, by determining the _m ^inp, Even with learning using a small number of learning models, a model suitable for the speaker can be obtained.

[Brief description of drawings]

FIG. 1 is a schematic configuration diagram of an HMM learning device according to the present invention.

FIG. 2 is a configuration diagram centering on a speaker adaptation unit of a learning device for an HMM according to the present invention.

FIG. 3 is a flowchart relating to a learning process and a speaker adaptation process according to the present invention.

FIG. 4 is a conceptual diagram for performing word recognition based on a conventional phoneme HMM.

FIG. 5: Input voices “ba (ba)” and “bu (b)
u) ”is a voice pattern showing the relationship between power and time.

FIG. 6 is a segment division diagram of phoneme / b /, and a diagram in which each segment is approximated by a Gaussian distribution.

FIG. 7 is a schematic configuration diagram of a conventional HMM learning device based on speaker adaptation of an HMM, and a speech recognition device using this learning device.

FIG. 8 is a diagram showing a correspondence before and after re-learning of an average vector of an initial model.

FIG. 9 is a conceptual diagram when performing a smoothing process of a moving vector.

FIG. 10 is a diagram showing a comparison result of recognition rates of respective representative speaker selection methods in the present invention.

[Explanation of symbols] 1 ... Voice analysis unit 2 ... Initial model storage 3 ・ ・ Learning department 4 ・ ・ Speaker adaptation department 5 ... ・ Adapted model storage 10a ... Speaker subspace movement vector calculation unit of input speaker 10b .. Speaker subspace movement vector storage unit of input speaker 10c ... Representative speaker selection unit 10d ... Model building unit after speaker adaptation 11 ··· Speaker subspace movement vector calculator of representative speaker 12 ··· Speaker subspace movement vector storage unit for representative speaker 13 ··· After adaptation model creation unit 14 ... Model storage after adaptation

Continuation of front page (56) References Okura, Onishi, Iida, Speaker adaptation of unspecified speaker HMM based on speaker space movement vector of multiple representative speakers, IEICE Technical Report [Speech], Japan , June 16, 1994, Vol. 94, No. 90, SP94-21, Pages 53-60 Okura, Onishi, Iida, Representative speaker selection method in speaker adaptation method based on speaker space movement vector, Acoustical Society of Japan 1994 Autumn Research Conference, Japan, October 31, 1994, 2-8-22, Pages 81-82 Okura, Onishi, Iida, Speaker adaptation of speaker-independent speaker model based on speaker space movement vector, ASJ 1994 Proceedings of Spring Meeting, Japan, March 23, 1994, 3-7-9, Pages 105-106 Miyanaga, Sagayama, Moving Vector Field Smoothing Examination of standard speaker selection method in speaker adaptation method , Proceedings of the 1994 Autumn Research Conference of the Acoustical Society of Japan, Japan, October 5, 1992, 2-5-2, Pages 121-122 Okura, Sugiyama, Sagayama, mixed continuous distribution H MM moving vector field Smoothing speaker adaptation method, IEICE Transactions DI , Japan, December 1993, Vol. J76-D-II, No. 12, Pages 2469-2476 (58) Fields investigated (Int.Cl. ⁷ , DB name) G10L 15/06 JISST file (JOIS)

Claims (2)

(57) [Claims] 1. A speaker subspace movement vector of an input speaker obtained from a small amount of input speaker learning audio material from speaker subspace movement vectors v _i , _s , _m ⁿ of a plurality of representative speakers. The speaker subspace movement vectors v _i , _s , _m ^spno of the representative speaker, which are close to v _i , _s , _m ^inp in distance, are selected, and the speaker subspace movement vectors v _i , _s , _{m of the} representative speaker are selected. An HMM learning device characterized by adapting an unspecified speaker HMM to an input speaker by modifying ^spno . 2. A voice analysis unit for analyzing characteristics of input voice.
(1), an initial model storage unit (2) for storing an initial model of the HMM, and a result of analyzing the voice of the input speaker in the voice analysis unit (1), and stored in the initial model storage unit (2) A learning unit (3) for learning the learned HMM, and an average vector μ _i , _s , _m of the HMM of the input speaker learned in the learning unit (3).
^Inp and the speaker subspace movement vector v _i of the input speaker, which is calculated using the difference vector obtained from the difference between the HMM average vectors μ _i , _s , and _m stored in the initial model storage unit (2), _s, speaker subspace movement vector calculating unit input speaker calculating the _m ^inp and (10a), the input speaker talking determined by speaker subspace moving vector calculation unit of the input speaker (10a) Speaker subspace movement vector v _i , _s , _m ^inp , and the speaker subspace movement vector v _i , _s , _m ⁿ of the representative speaker. The speaker subspace movement vector storage unit (12) of the representative speaker to be stored and the speaker subspace movement vector v of the input speaker stored in the speaker subspace movement vector storage unit (10b) of the input speaker. _i , _s ,
_{It is} obtained by the representative speaker selecting unit (10c) that selects the speaker subspace movement vectors v _i , _s , and _m ^spno of the representative speaker that are close to _m ^inp in distance, and the representative speaker selecting unit (10c). The representative speaker's speaker subspace movement vectors v _i , _s , _m ^spno , the input speaker's speaker subspace movement vectors v _i , _s , _m ^inp , and the initial model mean vectors μ _i , _s , _m . used, the mean vector mu _{i, s} after the speaker adaptation, speaker adaptation after model construction unit for determining the _m ^inp (10d)
If, HMM learning device characterized by comprising a mean vector mu _{i, s} after speaker adaptation, adaptive post-model storage unit for storing the _m ^inp (14), the.