JP3316355B2 - 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 learning apparatus for a Hidden Markov Model (hereinafter, referred to as "HMM"), which is a probabilistic model that approximately represents a statistical feature of speech by a distribution such as a Gaussian distribution.

[0002]

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

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

In general, when an HMM is created in phoneme units so that a word is formed by connecting smaller phonemes, for example, phonemes, word recognition for an arbitrary word is performed by linking the phoneme HMMs. Can be performed.

For example, recognition targets registered in the dictionary are “Uchisuke (U / CH / I / K / E / S / U)” and “Uchicho (U / CH / I / A / W / A / S / E) and “Uru (U / R / U)”, the phoneme HMM that needs to be created is “U / CH /
I / K / E / S / A / W / R ".

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 a word HMM close to the input speech (word) can be obtained as a probabilistic likelihood (probability).

In this way, by learning the speaker's voice information in advance and creating a phoneme HMM, it is possible to recognize even if the input voice is a word.

By the way, such a phoneme HMM is generally created using hundreds of learning words and the like. However, it is not realistic to make a user utter a few hundred words in consideration of the burden on the user. In order to avoid such a point, there is a speaker adaptation method as a method of tuning a phoneme HMM to a user's speech characteristics using a small number of learning words. Shu Ipp. 81-82 have announced speaker adaptation methods.

In general, in the HMM, speech features are represented by a probability density function such as a Gaussian distribution. For example, when speech is analyzed using a band filter of 16 channels, one Gaussian distribution is obtained for each channel. Here, by regarding the 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. Hereinafter, such a vector is referred to as an average vector. .

FIG. 6 shows that the present inventor has collected the above-mentioned papers of the Annual Meeting of the Acoustical Society of Japan, Autumn Research Conference, 1994, Ipp. 81
HM based on speaker adaptation of HMM announced at ~ 82
It is a schematic structure figure of the learning device of M.

In FIG. 6, reference numeral 1 denotes a voice analysis unit for analyzing characteristics of an input voice for each frequency band, and 2 denotes an initial model storage unit for storing an initial model of an HMM by learning. HMM of a specific speaker created by using the voices of speakers of the same speaker, or HMM of an unspecified speaker learned by using the voices of many speakers. As a specific learning method, a well-known forward backward algorithm, a learning rule based on Viterbi alignment, or the like may be used.

Reference numeral 3 denotes a learning unit for re-learning the above-described initial model using input speech.
It is assumed that only the average vector is learned from the parameters representing.

Reference numeral 4 denotes a speaker adaptation unit for making the HMM re-learned in the learning unit 3 into a more accurate model.

Reference numeral 5 denotes a speaker subspace movement vector calculation unit of the representative speaker for calculating the speaker subspace movement vector of the representative speaker used in the speaker adaptation unit 4.

Reference numeral 6 denotes a representative speaker's speaker subspace movement vector storage unit which stores the value calculated by the representative speaker's speaker subspace movement vector calculation unit 5.

Reference numeral 7 denotes a post-adaptive model generation unit that generates a post-adaptive model from the parameters of the HMM that has undergone speaker adaptation and other parameters.

Reference numeral 8 denotes a post-adaptive model storage unit that stores 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 of each unit will be described below.

The speaker subspace movement vector calculation section 5 of the representative speaker 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 average vector of the Gaussian distribution of the HMM after re-learning the initial model. A spatial movement vector can be determined.

Step 1: The initial model (λ) stored in the initial model storage unit 2 is used as the initial model of the phoneme HMM of each representative speaker.

Λ = {λ ₁ ,..., Λ _i ,..., Λ _I } where I indicates the number of the phoneme HMM. For example, when 39 phoneme HMMs are used, I = 39. Λ _i indicates the i-th phoneme HMM.

[0022] In addition, λ _i is _{λ i = {w i, s} , m, a i, s1,
_{s2, μ i, s, m} , σ i, s, represented by _m ^2}.

Note that w _i , _s , _m , μ _i , _s , _m , and σ _i , _s ,
_m ² is weight for the m-th Gaussian distribution of the s state of i-th phoneme HMM, mean vector, and a represents each vector of variance. a _{i, s1, s2} from the s1 state of i-th phoneme HMM represents the transition probability to the s2 state, for example when using a 33-dimensional vector as a feature quantity, mu
_i , _s , _m , σ _i , _s , and _m ² are 33-dimensional vectors.

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

Step 2: Learning is performed by connecting the HMMs of the representative speakers so as to correspond to the phoneme sequence of the input speech of the representative speakers. Learning w _{i, s, m} and μ _{i, s,} as the n-th representative speaker model performs _m only ^{_{λ i n = {w i n}} , s, m,
_{a i, s1, s2, μ} i n, s, m, σ i, s, obtaining _m ^2}. Here, n represents the number of the representative speaker, and n = 1, 2,.
N, for example, N when 10 representative speakers are used.
= 10.

[0026] Step 3: For each representative speaker, the difference t _i of the average value, _s, the _m ⁿ seek.

[0027] _{∀ i, s, m ∈Ω t} 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 , _m of the average vector μ _i , _s , _m included in λ. Represents

[0028] Step 4: As the number 1, obtaining representative speaker speaker subspace movement vector v _{i, s,} a _m ^n. 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 .

[0029]

(Equation 1)

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

[0031] On the other hand, t _{i, s,} the _m ⁿ may be used as the speaker subspace movement vector of the representative speaker.

[0032] In addition, learning _{{w i, s, m,} a i, s1, s2,
μ _i , _s , _m , σ _i , _s , _m ² }, at least μ _i , _s , _m
It is sufficient to learn to include. Of _{course, {w i, s, m} ,
_{a i, s 1, s2,} μ i, s, m, σ i, s, may be learned every _m ^2}.

A plurality of representatives calculated according to the above steps
Speaker subspace movement vector v _i,_s,_m ⁿThe representative story
Speaker subspace movement vector storage unit 6.

Next, a method of creating an adapted model of the input speaker will be described with reference to FIG.

The voice for model learning of the input speaker is analyzed by the voice analysis unit 1. Using the analysis result, the learning unit 3 learns the initial model.

Learning of the initial model: The initial model (λ) stored in the initial model storage unit 2 is stored in the phoneme HM of the input speaker.
Let M be the initial model.

[0037] _{λ = {λ 1, ···,} λ i, ···, λ I} Moreover, lambda _i is _{λ i = {w i, s} , m, a i, s1, s2, μ i, s ,
_m , σ _i , _s , _m ² }.

Next, learning is performed by connecting the HMMs of the input speakers so as to correspond to the phoneme sequences of the input speech of the input speakers. Learning is performed only on μ _i , _s , and _{m and} λ _i ^inp
_{= {W i, s, m} , a i, s1, s2, μ i inp, s, m, σ i, s,
_m ² ｝.

[0039] Next, based on the learning model by the learning unit 3, the input speaker of the speaker subspace movement vector v _{i, s,}
The _m ^inp obtained by the following steps. Here, inp represents the input speaker.

[0040] Step 1: calculate the difference t _i of the average value, _s, a _m ^inp.

[0041]

(Equation 2)

Step 2: An input speaker subspace movement vector v _i , _s , _m ^{inp is obtained according} to ^{Equation 3} .

[0043]

(Equation 3)

Here, E represents a set of subscripts of the average vector of the phoneme HMM corresponding to the phoneme appearing in the learning speech material.

The speaker subspace movement vector storage unit 4b of the input speaker is a speaker subspace movement vector calculation unit 4 of the input speaker.
input speaker subspace movement vector calculated in a v _{i, s, m}
Remember ^inp .

Branch probability of each component of phoneme HMM
And the input speaker subspace movement vector v_i,_s,_m
^inpSubspace movement vector of the representative speaker close to the distance
v_i,_s, _m ⁿNumber of the representative speaker (spno) with
Part of the representative speaker having the representative speaker number (spno) of
Spatial movement vector v_i,_s,_m ^spnoAccording to Equation 4
Is selected by the user selection section 4c.

[0047]

(Equation 4)

[0048] In the speaker adaptation after the model construction unit 4d, a representative of the representative speaker obtained in the speaker selection unit 4c speaker subspace movement vector ^v _{i, s, m} spno, of the input speaker speaker subspace movement vector v _i, with _{s, m} ^inp, and mean vector mu _{i, s} of the initial model, the _m, as the number 5, the mean vector mu _{i, s} after the speaker adaptation, seek _m ^adapt.

[0049]

(Equation 5)

The initial stored in adapted after model creation part 7, the mean vector mu _{i, s} after been Speaker Adaptation built in speaker adaptation after model construction unit 4d, _m ^inp, as well as the initial model storage unit 2 The weights w _i , _s , _m , for the Gaussian distribution of the model,
Transition probability a _{i, s1, s2} and variance vector σ _{i, s,} a weight w _i for Gaussian distribution of _m ² or input speaker, _{s, m} ^inp, transition probability a _{i, s1, s2} ^inp and variance vector sigma _{i, s, m} ^inp
Or representative speaker speaker space moving weight for a Gaussian distribution stored in the vector storage unit ^{_{6 w i, s, m spno}} , transition probability ^a _{i, s1, s2} spno and variance vector (σ _{i, s, m}
^{Use spno 2} to create a model after adaptation.

Finally, the post-adaptation model created by the post-adaptation model creation section 7 is stored in the post-adaptation model storage section 8.

[0052]

However, as described above, the following problems have been encountered in the preparation of a conventional HMM.

In the conventional HMM creation, the representative speaker closest to the input speaker's speaker subspace movement vector from the representative speakers stored in the speaker's speaker subspace movement vector storage section 6 is The speaker subspace movement vector is used for model learning, and in order to cover the speech characteristics of an unspecified number of input speakers, the speaker subspace movement vector of a number of representative speakers is used as the representative speaker. Has to be stored in the speaker subspace movement vector storage unit 6.

In other words, the use of the speaker subspace movement vectors of a large number of representative speakers causes an increase in storage capacity, and further, in order to calculate the speaker subspace movement vectors of a large number of representative speakers, There is a problem that it is necessary to collect voices of many people.

Also, the speaker subspace movement vector of the representative speaker is a vector with a small bias because learning is performed using speech that the recognition model can sufficiently learn. The speaker subspace movement vector of the input speaker obtained from the training voice depends on the features of the model learning voice. For example, if the learning voice is only the word “Akai”, only the features included in the learning utterance “Akai” will be learned.

That is, there is no speech feature that is not included in the learning utterance “Akai” such as “U” or “E”, and the speaker subspace movement vector is biased to the content of the learning word.

The speaker subspace movement vector of the representative speaker with less bias learned using such sufficient speech,
With the conventional representative speaker selection method based on the similarity between the input speaker's speaker subspace movement vectors depending on the characteristics of the training speech, it was difficult to perform highly accurate speaker selection.

[0058]

Therefore SUMMARY OF THE INVENTION The present invention has been made in view of the problems described above, a plurality of representative speakers of the speaker subspace movement vector v _{i, s,} any combination of the _m ⁿ most similar inner episode's speaker subspace movement vector v _i, which is calculated to correspond, _s, among the _m ^{inter (p),} speaker subspace movement vector v _i of the input speaker, _s, and _m ^inp sexual selects high speaker subspaces movement vector,該話's subspace movement vector ^{_{v i, s, m inter (}} spno) and the input speaker of the speaker subspace movement vector v _{i, s, m} ^inp and input mean vector mu _{i, s} of speaker HMM, characterized by learning HMM by using the _m ^inp.

Further, the present invention analyzes characteristics of an input voice.
Speech analysis unit (1) and initial memory for storing initial model of HMM
In the model storage unit (2) and the speech analysis unit (1),
Model storage unit using the results of analyzing the voice of the elderly (2)
A learning unit (3) for learning the HMM stored in the learning unit;
Average vector of HMM of input speaker trained in (3)
Le μ_i,_s,_m ^inpIs stored in the initial model storage unit (2).
HMM mean vector μ_i,_s,_mFrom the difference
Part of the input speaker calculated using the difference vector
Spatial movement vector v_i,_s,_m ^inpTalking input speaker to calculate
Speaker subspace movement vector calculation unit (4a), and the story of the input speaker
Input obtained by the user subspace movement vector calculator (4a)
Speaker subspace movement vector v_i,_s,_m ^inpMemorize
A speaker subspace movement vector storage unit (4b) for the input speaker,
Speaker subspace movement vector v of the representative speaker_i,_s,_m ⁿRemember
Speaker subspace movement vector storage unit for representative speaker (6)
And a speaker subspace movement vector storage unit for the representative speaker (6)
Story of N representative speakers (N is an integer of 1 or more) stored in
Subspace movement vector v_i,_s,_m ⁿFor speaker interpolation from
Interpolator selection for all combinations of P names
Selection section (9a) and the combination selected by the interpolated speaker candidate selection section (9a).
The speaker is interpolated based on the matching, and the speaker part of the interpolated speaker is empty.
Inter-movement vector v_i,_s,_m ^{inter (p)}Interpolator to calculate
Speaker subspace movement vector calculator (9b)
An interpolator that stores the speaker's speaker subspace movement vector
Speaker subspace movement vector storage unit (9c) and the input speaker
Input stored in speaker subspace movement vector storage unit (4b)
Speaker subspace movement vector v _i,_s,_m ^inpAnd similarity
Speaker subspace movement vector storage unit for high-interpolation speaker (9c)
Speaker subspace movement vector of the interpolated speaker stored in
v_i,_s,_m ^{inter (spno)}Representative speaker selection section (4c)
And the speaker of the interpolated speaker obtained by the representative speaker selection unit (4c).
Subspace movement vector v_i,_s,_m ^{inter (spno)}, Input speaker
Speaker subspace movement vector v_i,_s,_m ^inp, And the initial mode
Dell mean vector μ_i,_s,_mIs used to adjust the speaker
Mean vector μ_i,_s,_m ^adaptModel after speaker adaptation
And a construction unit (4d).

[0060] The present invention is representative speaker of the speaker subspace movement vector v _i of N name, _s, from _m ^n, high affinity speaker subspace movement vector v _i of the input speaker, _s, and _m ^inp Top n 'name (n' ≦ n) surrogate table speaker speaker subspace movement vector v _i of, _s, select _m ^n, selected representative speaker speaker subspace movement vector v _{i, s} speaker subspace motion vector inner episode who is calculated to correspond to a combination of _m ⁿ v _{i, s, m}
from the ^{inter (p),} speaker subspace movement vector v _i of the input speaker, _s, most similarity with _m ^inp selects high speaker subspaces movement vector,該話's subspace movement vector v _i , _S ,
_m ^{inter (spno)} and the speaker subspace movement vector v of the input speaker
_{i, s, m} ^inp and the average vector of the input speaker HMM mu _{i, s,} characterized by learning HMM by using the _m ^inp.

The present invention provides an audio component for analyzing the characteristics of an input audio.
Analyzer (1) and an initial model that stores the initial model of the HMM
The memory of the input speaker is stored in the storage unit (2) and the voice analysis unit (1).
Stored in the initial model storage unit (2) using the results of voice analysis
Learning unit (3) for learning the obtained HMM, and the learning unit (3)
Vector of HMM of input speaker trained by learning
μ_i,_s,_m ^inpAnd stored in the initial model storage unit (2)
HMM mean vector μ_i,_s,_mCalculated from the difference
Speaker partial sky of input speaker calculated using difference vector
Inter-movement vector v_i,_s,_m ^inpCalculate the input speaker's speaker
A subspace movement vector calculation unit (4a) and a speaker of the input speaker
Input story obtained by the subspace movement vector calculator (4a)
Speaker's subspace movement vector v_i,_s,_m ^inpMemorize
The storage unit (4b) for the speaker subspace movement vector of the input speaker
Speaker's speaker subspace movement vector v_i,_s,_m ⁿMemorize
A speaker subspace movement vector storage unit (6) for the representative speaker,
The above-mentioned input speaker's speaker subspace movement vector storage unit (4b)
Speaker subspace movement vector of stored input speaker
v_i,_s,_m ^inpAnd the representative speaker's speaker subspace movement vector v
_i,_s,_m ⁿA similarity calculation unit (9d-1) that calculates the similarity between,
Similarity based on the value calculated by the similarity calculator (9d-1)
Of upper interpolated speaker candidates that are sorted and stored in descending order
(9d-2) and the upper interpolated speaker candidate creation unit (9d-2).
N 'names (N' is an integer of 1 or more)
Speaker's speaker subspace movement vector v_i,_s,_m ⁿSpeaker from
And the interpolated speaker's subspace movement vector
v_i,_s,_m ^{inter (p)}Speaker subspace of the interpolator that computes
The motion vector calculator (9b) and the calculated speaker unit of the interpolated speaker
The speaker subspace of the interpolator that stores the spatial motion vector
The motion vector storage unit (9c) and the speaker subspace
The speaker part of the input speaker stored in the motion vector storage unit (4b)
Spatial movement vector v_i,_s,_m ^inpInterpolator with high similarity to
Of the speaker subspace movement vector storage unit (9c)
Interpolator speaker subspace movement vector v _i,_s,_m
^{inter (spno)}A representative speaker selection unit (4c) for selecting
The speaker subspace shift of the interpolated speaker obtained by the speaker selection unit (4c)
Motion vector v_i,_s,_m ^{inter (spno)}The speaker part of the input speaker
Spatial 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 ^adaptModel builder after speaker adaptation (4d)
And characterized in that:

Further, according to the present invention, the upper interpolated speaker candidate creating section (9
d-2), the upper N 'names (where N' is 1 or more)
Speaker subspace movement vector of the representative speaker of the above integer)
v_i,_s, _m ⁿFrom the speaker and the speaker subspace of the interpolated speaker
Movement vector v_i,_s,_m ^{inter (p)}When calculating
Talk from all combinations of the top N 'representative speakers
It interpolates the person.

[0063] The present invention is higher N 'name (N' in accordance with the stored order to the upper inside episode candidate creating unit (9d-2) is an integer of 1 or more) representatives speaker speaker subspace movement vector v _i of , _{s, m} ⁿ
Interpolating the speaker from the speaker subspace movement vector v _i of the inner episode's, _s, in the case of calculating the _m ^{inter (p),} talk sequentially with representative speakers from N 'name until the P-position It interpolates the person.

[0064] The present invention is learning representative speakers calculated using the utterance having the same content as the input talker speech that was used to calculate the input speaker of the speaker subspace movement vector v _{i, s, m} ^inp based the word dependent speaker subspace motion vector ζ _{i, s, m} ^n, speaker subspaces of the input speaker movement vector v _{i, s,} a similarity to _m ^inp, by performing a representative speaker selection It is characterized by.

[0065] The present invention is learning representative speakers calculated using the utterance having the same content as the input talker speech that was used to calculate the input speaker of the speaker subspace movement vector v _{i, s, m} ^inp perform word dependent speaker subspace motion vector ζ _{i, s,} and _m ^n, input speaker speaker subspace movement vector v _{i, s,} a representative speaker selection on the basis of similarity to _m ^inp, the selected representative speaker corresponding to the number of the representative speaker speaker subspace movement vector ^v _{i, s, m} spno an input speaker of the speaker subspace movement vector v _{i, s, m} ^inp, and initial model Mean vector μ _i , _s ,
_The feature is that the HMM is learned by using _m .

[0066] Further, the present invention was calculated using the utterance of the same content and audio input speakers used to calculate the speaker subspace of the input speaker movement vector v _{i, s, m} ^inp representative speaker make learning language dependent speaker subspace movement vector ζ _{i, s,} and _m ^n, the input speaker of the speaker subspace movement vector v _{i, s,} a representative speaker selected based on the similarity of the _m ^inp , the selected representative speaker representative speaker corresponding to the number of the speaker subspace movement vector v _{^i, s, m} spno learning language dependent speaker subspace motion vector zeta _{i, s,} and _m ^SPNO, input The speaker's speaker subspace movement vector v _i , _s , _m ^inp and the average vector μ of the initial model
HMM learning is performed by using _i , _s , and _m .

Also, the present invention provides a ^method of ^calculating the difference between the speaker subspace movement vector v _i , _s , _m ^spno of the representative speaker and the learning word dependent speaker subspace movement vector ζ _i , _s , _m ^spno , and the input speaker. The method is characterized in that an average vector μ _i , _s , _m ^adapt after speaker adaptation is obtained from a speaker subspace movement vector v _i , _s , _m ^inp and an average vector μ _i , _s , _{m of the} initial model.

According to the present invention, a speech analysis unit (1) for analyzing the characteristics of an input speech, an initial model storage unit (2) for storing an initial model of an HMM, and a speech of an input speaker in the speech analysis unit (1). A learning unit (3) that learns the HMM stored in the initial model storage unit (2) using the result of the analysis, and an average vector μ _i of the HMM of the input speaker trained in the learning unit (3). The speaker subspace movement vector of the input speaker calculated using the difference vector obtained from the difference between _s , _m ^inp and the HMM average vector μ _i , _s , _m stored in the initial model storage unit (2) v _{i, s,} speaker subspace movement vector calculating unit input speaker calculating the _m ^inp and (4a), the input spoken obtained by speaker subspace moving vector calculation unit of the input speaker (4a) 's speaker subspace movement vector v _{i, s,} speaker subspaces mobile base input speaker for storing _m ^inp Torr storage unit and (4b), representative speaker speaker subspace movement vector v _{i, s,} speaker subspace motion vector storage unit of the representative speaker for storing _m ⁿ and (6),
Speaker subspace movement vector v _i of the input speaker, _s, learning language dependent speaker subspace of representatives speaker that was created by using the voice of the representative speaker of the same content and audio input speaker of calculation of the _m ^inp movement vector zeta _{i, s,} learning language dependent speaker subspace motion vector storage unit of the representative speaker for storing _m ⁿ (11), stored in the input speaker of the speaker subspace motion vector storage unit (4b) by speaker subspace movement vector of the input speaker v _{i, s, m} ^inp highest similarity learning language dependent speaker subspace motion vector ζ _{i, s,
m} Representative speaker selection section that selects a representative speaker number (SPNO) with ⁿ and (4c), said selected representative speaker corresponding to the number of the representative speaker speaker subspace movement vector v _{i, s} , _m ^spno
And the input speaker's speaker subspace movement vector v _i , _s , _m ^inp ,
And a speaker-adaptive model constructing unit (4d) for ^obtaining the speaker-adapted average vector μ _i , _s , _m ^adapt using the initial model average vector μ _i , _s , _m. And

According to the present invention, a speech analysis unit (1) for analyzing the characteristics of an input speech, an initial model storage unit (2) for storing an initial model of an HMM, and a speech of an input speaker in the speech analysis unit (1). A learning unit (3) that learns the HMM stored in the initial model storage unit (2) using the result of the analysis, and an average vector μ _i of the HMM of the input speaker trained in the learning unit (3). The speaker subspace movement vector of the input speaker calculated using the difference vector obtained from the difference between _s , _m ^inp and the HMM average vector μ _i , _s , _m stored in the initial model storage unit (2) v _{i, s,} speaker subspace movement vector calculating unit input speaker calculating the _m ^inp and (4a), the input spoken obtained by speaker subspace moving vector calculation unit of the input speaker (4a) 's speaker subspace movement vector v _{i, s,} speaker subspaces mobile base input speaker for storing _m ^inp Torr storage unit and (4b), representative speaker speaker subspace movement vector v _{i, s,} speaker subspace motion vector storage unit of the representative speaker for storing _m ⁿ and (6),
Speaker subspace movement vector v _i of the input speaker, _s, learning language dependent speaker subspace of representatives speaker that was created by using the voice of the representative speaker of the same content and audio input speaker of calculation of the _m ^inp movement vector zeta _{i, s,} learning language dependent speaker subspace motion vector storage unit of the representative speaker for storing _m ⁿ (11), stored in the input speaker of the speaker subspace motion vector storage unit (4b) by speaker subspace movement vector of the input speaker v _{i, s, m} ^inp highest similarity learning language dependent speaker subspace motion vector ζ _{i, s,
m} Representative speaker selection section that selects a representative speaker number (SPNO) with ⁿ and (4c), said selected representative speaker corresponding to the number of the representative speaker speaker subspace movement vector v _{i, s} , _m ^spno
And learning word dependent speaker subspace movement vector ζ _i , _s , _m ^spno
And the speaker subspace movement vector v _i , _s , _m ^{inp of} the input speaker
And a second speaker-adapted model construction unit (4e) for ^obtaining the speaker-adapted average vector μ _i , _s , _m ^adapt using the initial model average vector μ _i , _s , _m It is characterized by the following.

The present invention provides a second speaker-adapted model construction unit (4e).
, The speaker subspace movement vectors v _i , _s ,
_m ^spno and learning word dependent speaker subspace movement vector ζ _i , _s , _m
The difference between ^spno and the speaker subspace movement vector v of the input speaker
_The feature is that an average vector μ _i , _s , _m ^adapt after speaker adaptation is obtained from _i , _s , _m ^inp and an average vector μ _i , _s , _{m of the} initial model.

[0071]

DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described with reference to FIGS.

In the configuration diagrams of FIGS. 1 to 5, the conventional HMM
The same components as those of the learning device are denoted by the same reference numerals.

FIG. 1 is a schematic configuration diagram of the first embodiment of the present invention.

The first point of the configuration of the HMM learning apparatus according to the first embodiment of the present invention which is different from that of the prior art is that the speaker of the interpolation speaker connected to the speaker subspace movement vector storage unit 6 of the representative speaker is used. That is, a user partial space movement vector creating unit 9 is provided.

As will be described later (see FIG. 2), the interpolated speaker's speaker subspace movement vector creating section 9 includes an interpolated speaker candidate selecting section 9a, and an interpolated speaker's speaker subspace movement vector calculating section 9b. , And an interpolated speaker subspace movement vector storage unit 9c.

The second difference of the configuration of the HMM learning apparatus of the first embodiment of the present invention from the conventional one is that the speaker subspace movement vector storage unit 6 of the representative speaker, the speaker adaptation unit 4 and the adaptation unit The connection of the interpolated speaker's speaker subspace movement vector creation unit 9, the speaker adaptation unit 4, and the post-adaptation model creation unit 7 is provided instead of the connection of the post-model creation unit 7.

FIG. 2 is a block diagram of the HMM learning apparatus according to the first embodiment of the present invention.

Hereinafter, the function of the speaker subspace movement vector creating unit 9 of the interpolated speaker, which is a typical configuration requirement in the first embodiment of the present invention, will be described in detail.

The interpolated speaker candidate selection unit 9a calculates all combinations of P names from the speaker subspace movement vectors of N representative speakers stored in the speaker subspace movement vector storage unit 6 of the representative speaker. the making, representative speaker of the speaker subspace motion vector v _i corresponding to this, _{s, m} ^n, the weights w _{i, s, m} ^n,
It sends mean vector μ _{i, s, m} ^n, the transition probability a _{i, s1, s2} ^n, and dispersion vector ^{_{(σ i, s, m 2}} ) to ⁿ of the inner episode's speaker subspace movement vector calculating unit 9b . Inner episode's speaker subspace movement vector calculating unit 9b in sent becoming a representative speaker speaker subspace movement vector v _{i, s,} speaker subspace motion vector inner episodic's using _m ⁿ v _i , _S , _m ^{inter (p)}
Is calculated.

The above processing will be described in detail for a case where five representative speakers are (a, b, c, d, e) (N = 5).

First, the interpolated speaker candidate selecting section 9a sets P = 1
In the case of (a), (a), (b), (c), (d), and (e) are sequentially sent to the speaker subspace movement vector calculation unit 9b of the interpolation speaker.

Next, assuming that P = 2, ten combinations of representative speakers selecting two out of five are sequentially sent to the speaker subspace movement vector calculator 9b of the interpolated speaker. Where 1
The 0 ways are (a, b), (a, c), (a, d),
(A, e), (b, c), (b, d), (b, d),
(C, d) (c, e) and (d, e).

Next, 10 combinations of P = 3, 5 combinations of P = 4, and 1 combinations of P = 5
The combinations of the representative speakers are sequentially sent to the speaker subspace movement vector calculator 9b of the interpolation speaker.

In the case of this example, 31 ways (5 + 10 + 10 +
5 + 1) representative speaker speaker subspace movement vector v _i corresponding to a combination of, _{s, m} ⁿ is sent to the speaker subspace movement vector calculating portion 9b of sequential internal episode's.

[0085] the episode's speaker subspace movement vector calculating unit 9b in inner episode candidate speaker subspace of sequentially sent the incoming representative speaker from the selection unit 9a movement vector v _{i, s,} the weight of the _m ⁿ A speaker subspace movement vector v _i , _s , _m ^{inter (p)} of the interpolated speaker is calculated according to Equation 6 as a weighted average.

At the same time, the speaker's subspace movement of the interpolated speaker
The vector calculator 9b calculates the variance vector of the interpolator
(Σ_i,_s,_m ^{inter (p)}), Weight w_i,_s,_m ^{inter (p)}, Trans
Transfer probability a_i, _s1,_s2 ^{inter (p)}Is calculated according to Equation 7.

In the case of this example, the speaker subspace movement vectors v _i , _s , and _m ^{inter (p ) of the} 31 interpolated speakers are calculated.

[0088]

(Equation 6)

[0089]

(Equation 7)

Here, Φ is an element in the 31 combinations in this example. Further, WGHT (n) is the weight for a representative speaker speaker subspace movement vector v _{i, s, m} ^n.

The calculated interpolated speaker's speaker subspace movement vector v _i , _s , _m ^{inter (p)} is stored in the interpolated speaker's speaker subspace movement vector storage unit 9c.

The interpolated speaker's speaker subspace movement vector v _i , _s , _m ^{inter (p)} stored in the interpolated speaker's speaker subspace movement vector storage section 9c is the speaker subspace of the representative speaker. movement vector v _{i, s,} is of _m ⁿ the same quality, in order to correspond to the case of increasing the interpolation number representative speaker, by the subsequent process of performing the same processing as the conventional example, after adaptation A model can be created.

In the case of this example, 31 pseudo representative speakers (the above interpolated speakers) can be created from five representative speakers.

For example, the representative speaker selecting section 4c calculates the speaker subspace movement vector v _i , _s , _m ^{inter (p)} of the interpolation speaker and the speaker subspace movement vector v _i , _s , _m of the input speaker. the distance between the ^inp calculated, input speaker subspace movement vector v _{i, s, m} ^inp
The distance to move subspace of the inner episode's close vector v _{i, s, m} episode's number among with ^{inter (p) (spn}
o), and an inner episode's speaker subspace movement vector v _i having inner episode's number ^{_{(spno), s, m inter}} (spno)
Is sent to the model adaptation unit 4d after the speaker adaptation, and the average vectors μ _i , _s and _m ^adapt after the speaker adaptation are obtained.

The post-adaptation model creation unit 7 calculates the average vector μ _i , _s , _m ^adapt after speaker adaptation, the variance vector (σ _i , _s , _m ^{inter (spno)} ) of the ^{spno-number interpolated speaker} , and the weight w _i , _s , _m
A model after adaptation is created based on ^{inter (spno)} and transition probabilities a _i , _s1 and _s2 ^{inter (spno)} .

FIG. 3 is a block diagram of an embodiment of the second HMM learning apparatus of the present invention.

The first difference of the configuration of the HMM learning apparatus according to the second embodiment of the present invention from the conventional one is that the speaker of the interpolation speaker connected to the speaker subspace movement vector storage unit 6 of the representative speaker is used. That is, a user partial space movement vector creating unit 9 is provided.

The interpolated speaker's speaker subspace movement vector creating section 9 includes a second interpolated speaker candidate selecting section 9d, an interpolated speaker's speaker subspace movement vector calculating section 9b, and an interpolated speaker's speaker section. It comprises a space movement vector storage unit 9c.

The second interpolated speaker candidate selecting section 9d includes a similarity calculating section 9d-1 and an upper interpolated speaker candidate creating section 9d-.
2

The second point of the configuration of the HMM learning apparatus according to the second embodiment of the present invention, which is different from the conventional one, is that the representative subspace movement vector storage unit 6 and the representative speaker selection unit 4 of the representative speaker are provided.
In place of the connection between c and the adapted model creating unit 7, a connection between the interpolated speaker's speaker subspace movement vector creating unit 9, the representative speaker selecting unit 4c and the adapted model creating unit 7 is provided.

The third point of the HMM learning apparatus according to the second embodiment of the present invention, which is different from the conventional HMM learning apparatus, is that the speaker subspace movement vector storage unit 4b of the input speaker and the interpolated speaker candidate selection unit 9d. Is provided.

The structure of the HMM learning apparatus according to the second embodiment of the present invention is different from that of the first embodiment in that the interpolated speaker candidate selecting section 9a is replaced with the second interpolated speaker candidate selecting section 9a.
d is provided.

Hereinafter, the HMM according to the second embodiment of the present invention will be described.
The function of the speaker subspace movement vector creating unit 9 of the interpolated speaker, which is a typical configuration requirement of the learning device of the first embodiment, will be described in detail.

[0104] In similarity calculator 9d-1, a representative speaker speaker subspace stored in the motion vector storage unit 6 representatives speaker speaker subspace movement vector v _{i, s, m} ⁿ as input speaker speaker subspaces motion vector storage unit 4b in the stored input speaker of the speaker subspace movement vector v _{i, s,} calculates the distance between the _m ^inp.

The upper interpolated speaker candidate creating section 9d-2 selects the upper N ′ names (N ′ ≦ N) having a smaller distance value among the distances obtained by the similarity calculator 9d-1, and targets the N ′ names. the similar to the one embodiment make any combination of P name, representative speaker speaker subspace movement vector v _i corresponding thereto, _{s, m} ^n, the weights w _{i, s, m} ^n, the average vector μ _{i, s, m} ^n, and sends the transition probability a _{i, s1, s2} ^n, and dispersion vector ^{_{(σ i, s, m 2}} ) to ⁿ of the inner episode's speaker subspace movement vector calculating unit 9b.

[0106] For example N '= If it is 5, the first embodiment as well as representatives speaker corresponding to a combination of 31 kinds speaker subspace movement vector v _{i, s, m} ⁿ are sequentially inner episode's sent to the speaker subspace movement vector calculating unit 9b, speaker subspace movement vector calculating portion 9b of the inner episode who, speaker subspace motion vector inner episode's for each combination v _{i, s, m} ^inter Calculate ^(p) .

At the same time, the interpolated speaker's speaker subspace movement vector calculation section 9b calculates the interpolated speaker's variance vector (σ _i , _s , _m ^{inter (p)} ) and weights w _i , _s , _m ^{inter (p ),} and calculates in accordance with transition probability a _{i, s1, s2} ^inter several 7 ^(p).

By selecting N ′ names from N names, the speaker subspace movement vectors v _i , _s , and _m ^{inter (p) of the} interpolated speaker are selected.
This has the effect of reducing the amount of processing.

The upper interpolated speaker candidate creating section 9d-2
May be configured to sequentially send candidates from the N ′ name to the top P order to the speaker subspace movement vector calculation unit 9b of the interpolated speaker.

Here, to sequentially send the candidates up to the top P means the following processing.

An example in the case where N '= 5 and P = 5 will be described below.

The top five candidates are (a, b, c, d, e)
, The speaker subspace movement vector of (a) is first calculated by the speaker subspace movement vector calculation unit 9 of the interpolation speaker.
Send to b. Next, (a, b), then (a, b, c), then (a, b, c, d), and then (a, b, c, d, e) speaker subspace motion vector Are sequentially transmitted to the interpolator's speaker subspace movement vector calculation unit 9b.

[0113] In this example, speaker subspace movement vector calculating portion 9b of the inner episode who is sequentially sent, the representative speaker speaker subspace movement vector v _{i, s,} of five different related _m ⁿ each speaker subspace of the inner episode's five types corresponding motion vector to the combination v _{i, s,} calculating the ^{_m i} _nter ^(p).

At the same time, the interpolated speaker's speaker subspace movement vector calculator 9b calculates the interpolated speaker's variance vector (σ _i , _s , _m ^{inter (p)} ) and weights w _i , _s , _m ^{inter (p ),} and calculates in accordance with transition probability a _{i, s1, s2} ^inter several 7 ^(p).

The calculated speaker subspace movement vector v _i , _s , _m ^{inter (p)} of the interpolation speaker is stored in the interpolation speaker's speaker subspace movement vector storage section 9c.

The representative speaker selecting section 4c is a speaker section of the interpolated speaker.
Spatial motion vector v_i,_s,_m ^{inter (p)}And the story of the input speaker
Subspace movement vector v_i,_s,_m ^inpCalculate distance with
And the input speaker subspace movement vector v_i,_s,_m ^inpAnd distance
Subspace movement vector v_i,_s,_m
^{inter (p)}And the number of the interpolator (spno) with
The speaker part of the interpolator with the number of the interrogator (spno)
Spatial movement vector v_i,_s, _m ^{inter (spno)}After speaker adaptation
The average vector μ after transmission to the Dell construction unit 4d and speaker adaptation
_i,_s,_m ^adaptAsk for.

The post-adaptive model creation unit 7 calculates the average vector μ _i , _s , _m ^adapt after speaker adaptation, the variance vector (σ _i , _s , _m ^{inter (spno)} ) of the ^{spno-number interpolated speaker} , and the weight w. _i , _s , _m
A model after adaptation is created based on ^{inter (spno)} and transition probabilities a _i , _s1 and _s2 ^{inter (spno)} .

FIG. 4 is a block diagram of an embodiment of the third HMM learning apparatus of the present invention.

The first difference of the configuration of the HMM learning apparatus of the third embodiment of the present invention from the conventional one is that it is connected to the representative speaker selection section 4c and the learning word dependent speaker subspace of the representative speaker. The difference is that a movement vector storage unit 11 and a learning word dependent speaker subspace movement vector calculation unit 10 of a representative speaker connected thereto are provided.

The second difference of the configuration of the HMM learning device of the third embodiment of the present invention from the conventional one is that the representative speaker number (spn) selected by the representative speaker selector 4c is used.
o) is sent to the speaker subspace movement vector storage unit 6 of the representative speaker.

Hereinafter, the speaker subspace movement vector calculating unit 10 which depends on the representative speaker's learning word and the representative speaker's learning word dependent speaker, which are representative components of the third embodiment of the present invention, will be described. This embodiment will be described in detail focusing on the subspace movement vector storage unit 11.

A learning word dependent speaker subspace movement pattern of a representative speaker
The function of the vector calculator 10 is that the input speaker is a learning word dependent speaker.
Basically the same function as the subspace movement vector calculation unit 4a
Having. The difference is that the speaker part of the input speaker depends on the adaptive word.
In the inter-movement vector calculation unit 4a, analysis is performed by the voice analysis unit 1.
Speaker subspace using input speaker's speech
Movement vector v_i,_s,_m ^inpTo calculate
Speaker's learning word dependent speaker subspace movement vector calculation unit 1
In the case of 0, the result of analyzing the voice of each representative speaker is used for the nth
Representative speaker's learning word dependent speaker subspace
Movement vector _i,_s,_m ⁿIs calculated. However
And the learning word dependent speaker subspace movement vector of the representative speaker ζ
_i,_s,_m ⁿThe voice of the representative speaker used to calculate
It is assumed that the content is the same as the voice of the input speaker.

The representative-speaker learning-word dependent speaker subspace movement vector storage unit 11 stores the representative-speaker learning-word dependent speech calculated by the representative-speaker learning-word dependent speaker subspace movement vector calculator 10. 's subspace motion vector ζ _{i, s,} stores the _m ^n.

[0124] In the representative speaker selection unit 4c, input speaker subspace movement vector v _{i, s, m} ^inp and distance to close representative speaker of learning language dependent speaker subspace movement vector ζ _{i, s, m} ⁿ Is selected and the representative speaker number (spno) is sent to the speaker subspace movement vector storage unit 6 of the representative speaker. Representative speaker of the speaker subspace motion vector storage unit of the representative speaker corresponding to 6 in a selected number of the representative speaker speaker subspace movement vector v _{i, s, m}
^{The spno} is ^transmitted to the speaker-adapted model 4d.

[0125] In the speaker adaptation after the model construction unit 4d, a representative of the speaker has been selected by the selection unit 4c representative speaker speaker subspace movement vector ^v _{i, s, m} spno, of the input speaker speaker subspace Using the motion vectors v _i , _s , _m ^inp and the average vector μ _i , _s , _m of the initial model, the average vector μ _i ,
^Find _s and _m ^adapt . Subsequent processing is the same as that of the conventional example, and a description thereof will be omitted.

FIG. 5 is a configuration diagram of an HMM learning apparatus according to a fourth embodiment of the present invention.

The first and second points in which the configuration of the HMM learning apparatus according to the fourth embodiment of the present invention is different from the conventional one are as follows.
This is the same as the third embodiment of the present invention.

The HMM learning apparatus according to the fourth embodiment of the present invention is significantly different from that of the third embodiment in that the post-speaker adaptation model constructing unit 4d is replaced by the second speaker adaptation model constructing unit. 4e.

[0129] The present In a fourth embodiment, the representative speaker selection section 4c input speaker subspace movement vector v _i in, _s, learning language dependent speaker subspaces movement of _m ^inp and distance to close the representative speaker until the process of selecting a vector ζ _{i, s,} representatives speaker numbers with _m ⁿ (SPNO) is identical to the third embodiment.

However, the representative speaker selection unit 4c sends the representative speaker number (spno) to the representative speaker's speaker subspace movement vector storage unit 6 and sends it to the second speaker adaptation model building unit 4e. The learning word dependent speaker subspace movement vector ζ _i , _s , _m ^spno of the representative speaker is transmitted.

[0131] In the second speaker adaptation after the model construction unit 4d, a representative speaker selection unit 4c speaker subspace movement vector of the selected representative speaker at ^v _{i, s, m} spno, representative speaker of learning language Using the dependent speaker subspace movement vector ζ _i , _s , _m ^spno , the input speaker's speaker subspace movement vector v _i , _s , _m ^inp , and the initial model average vector μ _i , _s , _m , According to 8, the average vector μ _i , _s , _m ^adapt after speaker adaptation is obtained.

[0132]

(Equation 8)

The subsequent processing is the same as that of the conventional example, and the description is omitted.

[0134]

As apparent from the above description the present invention is representative speaker speaker representative speaker speaker subspace movement vector subspace motion vector storage unit 6 N names stored in the v _i , _s, speaker subspace motion vector inner episode's from _m ⁿ v _{i, s,} by creating a _m ^{inter (p),} there is effect that artificially increasing the number of representative speakers, representatives speaker The number of representative speakers stored in the speaker subspace movement vector storage unit 6 can be reduced, and the storage capacity can be reduced.

Further, since the number of representative speakers is small, the number of people for calculating the speaker subspace movement vector of the representative speaker is small, and the speaker subspace movement vector of the representative speaker is created. This has the effect of reducing the amount of voice collection required to perform the operation.

[0136] The input speaker of the speaker subspace movement vector v _i obtained by calculating from the representative speaker's voice having the same content as the training speech input speaker, _s, the same bias as _m ^inp representative speaker of learning language dependent speaker subspace motion vector ζ _i have, _{s, m} ⁿ and, of input speaker speaker subspace movement vector v
order to carry out _{i, s,} a representative speaker selected by the similarity with the _m ^inp,
Highly accurate speaker selection can be performed in consideration of bias due to characteristics unique to the learning voice.

[0137] The representative speaker speaker subspaces motion vector storage unit 6 in the speaker subspace movement vector v _i representative speaker N names stored, _s, speaker inner episode's from _m ⁿ By creating the subspace movement vectors v _i , _s , and _m ^{inter (p)} , the number of representative speakers is increased in a pseudo manner, and highly accurate speaker selection is performed in consideration of the bias due to characteristics unique to the learning speech. As a result, highly accurate HMM learning becomes possible.

[Brief description of the drawings]

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

FIG. 2 is a configuration diagram of an HMM learning device according to the first embodiment of the present invention.

FIG. 3 is a configuration diagram of an HMM learning device according to a second embodiment of the present invention.

FIG. 4 is a configuration diagram of an HMM learning device according to a third embodiment of the present invention.

FIG. 5 is a configuration diagram of an HMM learning device according to a fourth embodiment of the present invention.

FIG. 6 is a conventional HMM learning device based on HMM speaker adaptation.

FIG. 7 is a detailed configuration diagram of a conventional HMM learning device based on HMM speaker adaptation.

[Explanation of symbols]

1. Voice analysis unit 2. Initial model storage unit 3. Learning unit 4. ... Speaker adaptation unit 4a ... Speaker subspace movement vector calculation unit of input speaker 4b ... Speaker subspace movement vector storage unit of input speaker 4c ········· Representative speaker selection unit 4d ·············································································· ······················································································································································ 記憶····· Adapted model creating unit 8 ········ Adapted model storage unit 9 ··················································································································· ... Interpolated speaker candidate selection unit 9b... Interpolated speaker subspace movement vector calculation unit 9c... Interpolated speaker subspace movement vector storage unit 9d ······· Second interpolated speaker candidate selection unit 9d-1 ····· Similarity calculation unit 9d-2 ····· Higher interpolated speaker candidate creation unit 10 ··········· Speaker learning word dependent speaker subspace movement vector calculation unit 11.... Representative speaker learning word dependent speaker subspace movement vector storage unit

Claims (2)

(57) [Claims] 1. A plurality of representative speakers of the speaker subspace movement vector v _{i, s, m} ⁿ speaker subspace movement vector v _i of the inner episode who is calculated to correspond to any combination of, _{s, m}
from the ^{inter (p),} speaker subspace movement vector v _i of the input speaker, _s, most similarity with _m ^inp selects high speaker subspaces movement vector,該話's subspace movement vector v _i , _S ,
_m ^{inter (spno)} and the speaker subspace movement vector v of the input speaker
_{i, s, m} ^inp and mean vector mu _{i, s} of the input speaker HMM, HMM learning device characterized by learning the HMM by using _m ^inp. 2. A high representative speaker speaker subspace movement vector v _i of N name, _s, the _m ^n, speaker subspace movement vector v _i of the input speaker, _s, and _m ^inp similarity higher N 'name (N' ≦ N)
The surrogate table speaker speaker subspace movement vector v _{i, s,} select _m ^n, speaker subspace movement vector v _i of the selected representative speaker, _s, calculated to correspond to a combination of _m ⁿ Of the interpolated speaker's speaker subspace movement vectors v _i , _s , _m ^{inter (p)} , the input speaker's speaker subspace movement vectors v _i , _s , _m
Select the most high similarity speaker subspace movement vector and ^inp,該話's subspace movement vector v _{i, s, m}
^{inter (spno)} and the input speaker of the speaker subspace movement vector v _{i, s, m} ^inp and the average vector of the input speaker of HMM μ
_{i, s,} HMM learning device characterized by learning the HMM by using _m ^inp.