JP3144203B2 - Vector quantizer - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION The present invention relates to a vector quantization (V
The present invention relates to an apparatus for speaker adaptation of a code book in pattern recognition and communication using Q (Vector Quantization) or speaker normalization of an input signal to be recognized and a signal to be transmitted.

[0002]

2. Description of the Related Art Vector quantization is widely used as a basic technique in high-efficiency encoding in transmission of speech signals and the like, and in pattern recognition such as speech recognition. Vector quantization is performed as follows.

A vector space to be handled is divided into M subspaces, and each subspace is labeled with a label (number) 1,..., M, and a subspace m corresponding to the label m is assigned.
(M = 1,..., M) representative vector (code vector)
Determines _{μ m, μ m (m =} 1, ..., M) using a codebook that stores the reference a form in m, converts the vector y label 1, ..., either of M is there. That is, when the distance between vectors u and v is d (u, v), y is a label

[0004]

(Equation 1)

[0005] The subspace is determined by clustering the training vector set. A well-known LBG algorithm is often used as a clustering method. In this case, the representative vector μ _m is the center of gravity or the average vector of the cluster m, and is also called the centroid of the cluster m.

Transmission of an audio signal using vector quantization is performed as follows. On the transmitting side, the PCM audio signal to be transmitted is divided into blocks every n samples, each block is regarded as an n-dimensional vector, and this is converted into a label sequence using the codebook. This will be described with reference to FIG. Reference numerals 2 and 3 denote buffer memories for alternately storing successive n samples. Reference numeral 1 denotes a switch for switching the buffer memories 2 and 3 so that the input n samples are stored alternately. A switch 4 alternately selects and outputs n samples from the buffer memories 2 and 3. 1 to 4 operate such that reading is performed from the other buffer memory while one buffer memory is performing writing. 5 is a code book,
An n-dimensional representative vector of each of the M clusters is stored in a form that can be searched for by a label. Reference numeral 6 denotes a comparison unit that compares the n-dimensional vectors stored in the buffer memories 2 and 3 with the M representative vectors stored in the codebook 5. Reference numeral 7 denotes a label selection unit for selecting a label corresponding to a representative vector most similar to the vector in the buffer memories 2 and 3 as a result of the comparison. The selected label is sent. That is, successive n samples are sequentially converted into labels,
This label is transmitted.

On the receiving side, the received label sequence is converted into a corresponding representative vector sequence using a codebook having the same configuration, and is converted back to a time waveform. Reference numeral 8 denotes a code vector reading unit, and 9 denotes a code book. The code book 9 has the same configuration as the code book 5. The n-dimensional code vector (representative vector) corresponding to the label received by using 8, 9 is read from the code book 9. Reference numerals 11 and 12 denote buffer memories for alternately storing n-element code vectors read from the code book 9, respectively.
The n-dimensional code vectors read from the code book are stored alternately. Reference numeral 10 denotes a switch for alternately distributing the code vectors read from the code book 9 to the buffer memories 11 and 12. Reference numeral 13 denotes a switch for alternately reading and outputting the contents of the buffer memories 11 and 12. The buffer memories 11 and 12 store the values approximated by the code vectors of the vectors of the buffer memories 2 and 3. Therefore, if this is read out serially for each element of the n-dimensional vector, a decoded signal can be obtained in a form approximating the transmission signal. When one of the buffer memories 11 and 12 is being read, the other is written to the other. Reading is alternately performed from the buffer memories 11 and 12 through the switch 13.

By doing so, for example, 1
When transmitting an audio signal whose sample is represented by 12 bits, the codebook size is M = 256, and the block length is n.
If = 8, the transmission bit rate is as follows. That is, the transmission amount per block is 12 × 8 = 96 [bits] when the PCM signal is transmitted as it is, but when the vector quantization is performed, the number of bits for distinguishing the label, that is,
log ₂ 256 = 8 [bits], and the transmission bit rate is 1/12. In this case, the vector y having n samples as components in each buffer memory is approximated (quantized) by the nearest centroid. Therefore, as the codebook size M increases, the quantization error decreases, but the number of bits required for encoding increases. In addition, the representative vector is obtained from the vector set prepared for learning as described above. In order to perform this with high accuracy, the larger the M becomes, the more the learning vector is required. Therefore, it is necessary to determine the codebook size according to the purpose by comprehensively taking into account the error accompanying the quantization, the transmission bit rate, the estimation accuracy of the representative vector, and the like.

The speech recognition apparatus converts an unknown speech signal into a sequence of acoustic feature vectors, and calculates the likelihood of each reference model stored in advance corresponding to each recognition category with respect to the acoustic feature vector sequence. Then, the likelihood is identified as the reference model having the maximum likelihood. FIG. 2 is a block diagram of a general speech recognition device using vector quantization. Reference numeral 20 denotes a feature extraction unit which converts an input speech signal into a feature vector. That is, for example, every 10 msec, it is converted into an n-dimensional feature vector by filter bank, LPC analysis, cepstrum analysis and the like.
Reference numeral 21 denotes a code book, which is clustered by a well-known clustering method from a set of feature vectors obtained in advance from the training speech in the same manner as described above, and labels each cluster. It is a memory of the centroid. Reference numeral 22 denotes a vector quantization unit, which includes the comparison unit 14 and the label selection unit 15 shown in FIG. Therefore, the feature extraction unit 2
The feature vector obtained at 0 is converted to the label of the centroid cluster closest to the feature vector with reference to the codebook 21. Reference numeral 23 denotes a reference model storage unit which stores reference models corresponding to each recognition unit. As recognition units, words, syllables, phonemes, and the like are often used. Reference numeral 24 denotes a reference unit, which is a reference model storage unit 23.
Vector quantization unit 22 for each of the reference models stored in
Calculate the likelihood for the label sequence obtained in the output of.
A determination unit 25 determines the recognition unit corresponding to the reference model having the maximum likelihood as a recognition result.

As reference models, there are a model having each recognition unit voice as a label sequence, a model having a so-called HMM (Hidden Markov Model) in which states and state transitions, and a degree of occurrence of a feature vector in each state are defined. Proposed.

The former is known as the SPLIT method, and includes a label sequence corresponding to an unknown input speech, a label sequence which is a reference model and a label sequence collated with each other, and a feature extraction unit 20 obtained from an unknown input. Is converted to a label without converting the distance vector to each centroid (a vector having the distance to each centroid in each frame as an element) or the similarity vector (the similarity to each centroid in each frame as an element In this method, the obtained distance (similarity) vector sequence is compared with the reference model.

The latter is a system which has recently become mainstream, and various improvements have been proposed, but are basically based on the following principle. The feature vector sequence for unknown input to be recognized _{Y = y 1, y 2,} ..., y T, any state sequence X = x ₁ length T generated from _{HMMλ, x 2, ..., x} T, the state The transition probability from i to state j is a _ij , the initial probability of state i,
When the probability of state i at t = 1 is π _i and the degree of occurrence of vector y _t in state i is ω _i (y _t ), the degree of occurrence of feature vector series Y from λ is (Equation 4)
Is shown as

[0013]

(Equation 2)

Or

[0015]

(Equation 3)

Or, taking the logarithm of both sides of (Equation 3),

[0017]

(Equation 4)

A state transition diagram of a frequently used model is represented as shown in FIG. However, in FIG.
Indicates that it corresponds to the recognition unit w. Assuming that this is HMM w, the reference model storage unit 23 in (FIG. 2) stores HMM 1, HMM 2,
.., HMM W is stored. At this time, the recognition results are L ₁ (Y | λ ^w ) and L ₂ (Y |
λ ^w ) and L ₃ (Y | λ ^w )

[0019]

(Equation 5)

## EQU1 ## However, in (Equation 5), i = 1 when (Equation 2) is used, and i when (Equation 3) is used.
= 2 and (Equation 4), i = 3.

Degree of occurrence of feature vector ω in state i
Depending on the definition of _i (y _t ), continuous HMM, discrete HM
M, FVQ type HMMs and the like exist. The present invention provides a discrete H
It relates to MM and FVQ HMMs.

In the discrete HMM, when b _im is the probability of occurrence of label m in state i,

[0023]

(Equation 6)

[0024] As an improvement of the discrete HMM, there is an HMM based on fuzzy vector quantization (FVQ HMM). In normal vector quantization, y
_{While t} is uniquely quantized to the representative vector of the cluster closest to it, fuzzy vector quantization is based on the degree of belonging of y _t to cluster m 0 ≦ u _tm ≦ 1, u _t1 + u
_t2 + ... + _utM = 1 is defined,

[0025]

(Equation 7)

Or

[0027]

(Equation 8)

Are defined as follows.

[0029]

Normally, a code book is obtained as an average value from uttered voices of various sentences, words, etc. of a large number of speakers. In the case of communication, the quality of the decoded signal deteriorates, and in the case of speech recognition, the recognition performance deteriorates. If a codebook is created for each speaker and the applied codebook is switched according to the speaker, the performance will be improved, but it will be necessary to collect a large amount of training data from one speaker, which is not practical.

[0030]

[Means for Solving the Problems] A standard codebook that stores a number of representative vectors in a feature vector space in a searchable form with corresponding labels, a learning vector storage unit that stores some learning vectors, Objective function calculating means for calculating an objective function defined as a function of a use vector, moving vector calculating means for calculating a moving vector, and adapting means for adding the moving vector to the representative vector to obtain a new representative vector. And converting the input vector into a label or a membership vector using the new representative vector as an element with the membership of the input vector with respect to each label when encoding the input vector. Means that the new representative vector is 2 and calculates to approximate the function extremum. A standard codebook that stores a number of representative vectors in a feature vector space in a searchable form with corresponding labels, a learning vector storage unit that stores some learning vectors, Function vector calculating means for calculating an objective function defined as a function of the input vector, moving vector calculating means for calculating the moving vector, and normalizing means for adding the moving vector to the input vector. In the conversion, the normalized input vector is obtained by adding the movement vector and the input vector, and the representative vector is used to convert the input vector into a label or a membership vector with the membership of each input vector as an element. Wherein the movement vector calculating means stores the motion vector in the standard codebook. And, when replacing the sum of the motion vector and the learning vector As a new learning vector, and calculates to approximate the objective function extremum.

[0031]

[Action]

1. Some representative vectors in the feature vector space are stored in a standard codebook in a searchable form with corresponding labels, and some learning vectors are stored in learning vector storage means. , An objective function defined as a function of the representative vector and the learning vector is calculated, a moving vector is calculated by a moving vector calculating means, and the moving vector is added to the representative vector by an adaptive means to obtain a new representative vector. And converting the input vector to be encoded by the vector quantization means into a label or a membership vector having the membership of each input vector as an element, using the new representative vector,
The calculation of the movement vector is to calculate the new representative vector so that the objective function approaches an extreme value with respect to the learning vector. 2. Some representative vectors in the feature vector space are stored in a standard codebook in a searchable form with corresponding labels, and some learning vectors are stored in learning vector storage means. Calculates an objective function defined as a function of the representative vector and the learning vector, calculates a motion vector by a motion vector calculation means, adds the motion vector to an input vector to be coded by a normalization means, and performs normalization. Converting the input vector normalized by the vector quantization means into a label or a membership vector using the representative vector as an element of the membership of the input vector with respect to each label. , The movement vector is calculated by the movement vector calculation means. Codebook to, and calculates to approximate the extrema of the objective function obtained by adding the motion vector to the learning vector as a new learning vector.

[0032]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention relates to speaker normalization or codebook adaptation. That is, in order to compensate for the above drawbacks, the present invention relates to a method of correcting an input vector according to a speaker or a method of correcting a representative vector of a codebook according to a speaker. This is executed without a teacher (it does not tell the system which word, sentence, etc. the speaker uttered) from the voice of (1).

The code book is created by clustering from a set of feature vectors obtained by uttering a number of speakers. Clustering methods include so-called hard clustering, in which each feature vector belongs to only one cluster, and fuzzy clustering, in which each feature vector belongs to each cluster according to the degree of belonging to each cluster. L for hard clustering method
There is an algorithm called a BG method, and a well-known method such as a fuzzy k-means method is used as a fuzzy clustering method. Although the present invention is used for both hard clustering and fuzzy clustering, hard clustering can be considered to be a special case of fuzzy clustering.

Fuzzy clustering is performed as follows. Y _1, y ₂ multiple speakers with a serial number feature vectors obtained by saying, ..., y _n, ..., and y _N. When the problem is, the cluster m of _{y n (= 1, ···,} M) u nm the degree of belonging to, the centroid vector of cluster m and μ _m, the objective function

[0035]

(Equation 9)

The centroid matrix V = [μ ₁ , μ ₂ ,..., Is minimized under the condition of u _n1 + un ₂ +... + _UnM = 1.
μ _M ] and the membership matrix U = [ _unm ]. This is performed by alternately repeating the operation of fixing one of V and U and minimizing the objective function J by the other for V and U. Specifically, V
Is fixed, U ′ is obtained as a solution for U of ∂J / ∂U = 0, and U is fixed to obtain V ′ as a solution for V of ∂J / ∂V = 0, U = U ′, V = The operation of changing V 'to new U and V is alternately repeated until convergence.
F is called fuzziness, F> 1 and
The ambiguity between clusters increases as F increases.

Fuzzy clustering is performed by the following steps. _{Here, d (y n, μ m} ) = (y n -μ m)
And ^T (y _n ^-μ _m). (Step 1-1) The number of clusters is M, the number of iterations is s = 0, the value of the objective function is J ⁽⁰⁾ = ∞, and the membership matrix U = [ _unm ]
Is appropriately given as an initial value U ⁽⁰⁾ . (Step 1-2) Set s = s + 1. (Step 1-3) Average vector μ of cluster m
_m ^(s) (m = 1,..., M) is obtained by the following equation.

[0038]

(Equation 10)

(Step 1-4) A matrix of the degree of belonging of each point to the cluster is calculated by the following equation.

[0040]

[Equation 11]

(Step 1-5) Calculation of Objective Function

[0042]

(Equation 12)

(Step 1-6) Termination condition

[0044]

(Equation 13)

When the condition is not satisfied (step 1-2)
What. When satisfied, end. Here, ε is a predetermined appropriately small positive number, and the smaller this value is, the higher the centroid estimation accuracy is, but the longer it takes to converge.

In the above steps, (Equation 10) is | J
^(s-1) / | by solving for ^{_{μ m (s-1) =}} 0 and μ _m ^(s-1), as (number 11) is undetermined multiplier of Lagrange the θ

[0047]

[Equation 14]

By solving u _nm ^(s-1) . If fuzziness F → 1 + 0, then 1 /
A (F-1) → ∞, when μ _m ^(s-1) is the highest close to _{y n, d (y n,} μ m (s-1)) <d (y n, μ h ( ^{s-1)) for h ≠} m d (y n, μ m (s-1)) = d (y n, μ h (s-1)) since it is for h = m, {d ( y n, ^{_{μ m (s-1))}} / d (y n, μ h (s-1))} 1 / (F-1) → 0 f
or h ≠ m {d (y n, μ m (s-1)) / d (y n, μ h (s-1))} 1 / (F-1) = 1 f
or h = m,

[0049]

(Equation 15)

That is, hard clustering is performed. Hard clustering in fuzzy clustering, when the label of the closest cluster y _n and L (n), is to define the ^{_{u nm (s) = δ L}} (n), m (s). Here, δ _ij is the Kronecker delta, δ _ij = 1 when i = j, and δ _ij = 0 when i ≠ j. Therefore, in the case of hard clustering, the above processing procedure is as follows.

First, the objective function is

[0052]

(Equation 16)

Is as follows. In this case, the clustering procedure is as follows. (Step 2-1) It is assumed that s = 0 and J ⁽⁰⁾ = ∞. (Step 2-2) s = s + 1 is set. (Step 2-3) Average vector μ _m ^(s) (m
= 1,..., M) by the following equation.

[0054]

[Equation 17]

(Step 2-4) The centroid closest to each point is calculated, and each point is clustered.

[0056]

(Equation 18)

(Step 2-5) Calculation of Objective Function

[0058]

[Equation 19]

(Step 2-6) Termination condition

[0060]

(Equation 20)

When the condition is not satisfied (step 2-2)
What. When satisfied, end. The codebook is created as described above, and the adaptation of the codebook thus created to the voice of speaker A is performed as follows.

The problem is that the centroid μ _m (m = 1,...,
M) to μ _m ′ so as to be most suitable for the voice of speaker A. The first embodiment according to the present invention, the transformation given by _{μ m '= μ m + h} m, and performs this by finding the optimal h _m from speech uttered by the speaker A. Specifically, feature numbers obtained from the voice uttered by speaker A for codebook adaptation are serially numbered, and y
^{When A} ₁ , y ^A ₂ , ..., y ^A _I ,

[0063]

(Equation 21)

H is appropriately reduced_mBy finding
Can be performed. As in the above example, d (y, μ) = (y−
μ)^T(y-μ), h_mBut
Desired. S is preset as the upper limit of the number of repetitions
Value. (Step 3-1) When the number of clusters is M,
The number of times is s = 0 and the value of the objective function is J⁽⁰⁾= ∞, h_m ⁽⁰⁾
= 0 (m = 1,..., M), and the membership degree matrix U = [u_nm]of
Initial value U ⁽⁰⁾Is given by the following equation.

[0065]

(Equation 22)

(Step 3-2) s = s + 1 is set. (Step 3-3) Movement vector h _m ^(s) (m = 1,...,
M) is calculated by the following equation.

[0067]

(Equation 23)

(Step 3-4) A matrix of the degree of belonging of each point (learning vector) to the cluster is calculated by the following equation.

[0069]

(Equation 24)

(Step 3-5) Calculation of Objective Function

[0071]

(Equation 25)

(Step 3-6) Termination condition

[0073]

(Equation 26)

If the condition is not satisfied (step 3-2)
What. When satisfied, end. Δ in (Step 3-6) is an appropriately small number, and is determined by how close the centroid of the codebook prepared as standard is to the speech input used for learning. δ
Is small and S is large, it approaches a codebook obtained by clustering only with the learning speech. When the number of learning voices is small, it is considered unfavorable that the distribution of the centroid is excessively biased toward the learning voices. Therefore, δ and S are appropriately selected according to the number of learning voices. Should.

[0075] When the voice for learning is small, but rather, a h _m in the objective function (number 21) m = 1, ···, it is better to be in common with respect to M. That is, the second embodiment according to the present invention is in this case, where h = h ₁ = h ₂ =... = H _M and the objective function is

[0076]

[Equation 27]

It is assumed that h is determined by the following steps. (Step 4-1) The number of clusters is M, the number of iterations is s = 0, the value of the objective function is J ⁽⁰⁾ = ∞, h ⁽⁰⁾ = 0, and the membership matrix U = [ _unm ] It gives the initial value U ⁽⁰⁾ by the following equation.

[0078]

[Equation 28]

(Step 4-2) s = s + 1. (Step 4-3) The movement vector h ^(s) is obtained by the following equation.

[0080]

(Equation 29)

(Step 4-4) The degree of membership of each point (learning vector) to the cluster is calculated by the following equation.

[0082]

[Equation 30]

(Step 4-5) Calculation of Objective Function

[0084]

(Equation 31)

(Step 4-6) Termination Condition

[0086]

(Equation 32)

When the condition is not satisfied (step 4-2)
What. When satisfied, end. Also in this case, the influence of the voice uttered for learning on the centroid correction amount can be adjusted by selecting δ and S.

FIG. 5 is a block diagram showing the structure of the first and second embodiments. In the case of the first embodiment, steps (3-1) to (3-6) are executed, and in the case of the second embodiment, (step 4-4) is executed.
-1) to (Step 4-6) are executed. 50 the learning vector y ^A ₁ for codebook creation, ..., the input terminal of the y ^A _N, 51 in the buffer memory, the _{^{y A 1, ..., y A}} N
Is stored. Reference numeral 54 denotes a standard code book, which stores code vectors created by a large number of speakers in a form that can be searched for by labels. 53 is a movement vector storage unit, 55
Is an adder, and the contents of the standard codebook 54 and the contents of the movement vector storage unit 55 are added by the adder 55. Reference numeral 52 denotes a movement vector calculation unit, which is based on the contents of the buffer memory 51 and the output of the adder 55 in the case of the first embodiment.
H _m (m = 1,..., M) is calculated according to 6), and the second
In the case of this embodiment, the above (Step 4-1) to (Step 4-6) are calculated. The calculated movement vector is stored in the movement vector storage unit 53. At the start of the repetitive calculation, the contents of the movement vector storage unit 53 are initialized to zero. According to this configuration, the contents of the movement vector storage unit 53 are rewritten each time by the updated movement vector obtained during the calculation. If the convergence condition of (Step 3-6) or (Step 4-6) is satisfied, a motion vector suitable for the speaker A is finally obtained in the motion vector storage unit 53. The motion vector obtained in this manner is added to the output of the standard godbook, so that a representative vector suitable for speaker A can be obtained.

FIG. 6 shows a case where the adaptation codebook 56 is inserted between the adder 55 and the movement vector calculation section 52. That is, with this configuration, it is apparent that the adapted codebook is finally obtained as a codebook suitable for the speaker A in the adapted codebook.

FIGS. 7 and 8 show an embodiment of the transmitting side of a communication device using the above principle.

FIG. 7 shows a method of speaker adaptation (FIG. 5).
This is a case where the one shown in FIG. Blocks 1, 2,
3, 4, 6, and 7 operate in the same manner as the blocks of the same number shown in FIG. The blocks 51 to 54 in FIG. 7 operate in the same manner as the blocks of the same number in FIG. 6, and are mostly used only for speaker adaptation. Each time the speaker changes, the motion vector indicating the deviation of the speaker from the standard codebook is learned and stored in the motion vector storage unit 53 according to the above description. In the case of the system shown in FIG. 1, the output of the switch 4 is compared with the contents of the codebook 5 by the comparing unit 6, but FIG.
Then, the output of the switch 4 and the output of the adder 55 are compared. It can be considered that the output of the adder 55 is obtained by correcting the standard codebook for the deviation of the speaker.

FIG. 8 shows a method of speaker adaptation (FIG. 6).
This is a case where the one shown in FIG. In this case, the adaptive codebook is inserted as described above. In this case, the comparator 6 compares the output of the switch 4 with the output of the adaptation codebook. That is, the adaptation codebook stores the representative vector obtained as a result of the correction for the speaker.

(FIG. 9) to (FIG. 12) show an embodiment of a receiver for reproducing the original sample sequence from the label sequence sent as described above.

In FIG. 9, the motion vectors according to the speaker are transmitted first, and they are stored in the motion vector storage unit in advance. Thereafter, the vector corresponding to the sent label is read from the standard codebook, and the read code vector is corrected by the adder 93 according to the contents of the moving vector storage unit. A similar process is performed to obtain a decoded signal.

FIG. 10 shows a case where the adaptive code book 101 is provided. That is, the addition output of the contents of the movement vector storage unit 92 and the contents of the standard codebook by the adder 93 is calculated in advance for all the code vectors and stored in the adaptation codebook. A book is used in place of the code book 9 in FIG.

FIG. 11 shows a case where the code book itself is transmitted from the transmitting side in advance, not the motion vector. That is, the contents of the adapted codebook created by the transmitter such as (FIG. 8) are transmitted and stored in the codebook 111. It goes without saying that the code book 81 corresponds to the code book 9 of FIG.

FIGS. 12 and 13 show an embodiment in which the above speaker adaptation method is applied to speech recognition.

FIG. 12 shows the case where the method of FIG. 5 is used, and 51 to 55 work in the same manner as in the case of FIG. Therefore, after the speaker adaptation, the output of the adder 55 is used instead of the codebook 21 of FIG.

FIG. 13 shows the case where the method of FIG. 6 is used, and 51 to 56 perform the same operation as in the case of FIG. Therefore, after the speaker adaptation, the adaptation codebook 56 is used instead of the codebook 21 of FIG.

The above description is based on the viewpoint of adapting the codebook to the speaker. However, in contrast to this, the method of adapting the speaker to the standard codebook, that is, a method of performing speaker normalization Is also conceivable. That is, (Equation 21) is

[0101]

[Equation 33]

[0102] because become, if Genzure the h _m from y ^A _i, can be considered to Fu say that normalized to the speaker in the code book. (Equation 33) corresponds to (FIG. 5) or (FIG. 6).
7) If the configurations of (a) and (b) are used, (Equation 34) is obtained corresponding to (Equation 33).

[0103]

(Equation 34)

(FIG. 14) is an embodiment of the transmitting side of the communication system based on vector quantization by speaker normalization according to the third embodiment, and uses (FIG. 5) or (FIG. 6). Is the case. 51 to 55 operate in the same manner as described above. In this case, the motion vector learned according to the above description is subtracted from the input vector, and vector quantization is performed using the standard codebook 54. A subtractor 131 subtracts a movement vector from an input vector.

(FIG. 15) is a receiver for the transmitter described in (FIG. 14), and converts a label sequence received using the standard codebook 91 into a code vector sequence.
The motion vector separately sent from the transmitting side is added to the code vector to obtain a decoded vector. An adder 141 performs this addition. Reference numeral 92 denotes a movement vector storage unit for storing a movement vector to be added by the adder 141, and this movement vector is transmitted from the transmitting side in advance when the speaker changes.

FIG. 16 shows an embodiment of the speech recognition apparatus based on vector quantization by speaker normalization according to the third embodiment. 51 to 55 operate in the same manner as described above. Also in this case, the motion vector learned according to the above description is subtracted from the input vector by the subtractor 131, and vector quantization is performed using the standard codebook 54.

The above-described vector quantization (FIG. 17)
It is clear that the transmission / reception device and the speech recognition device can be realized with substantially the same configuration when using the devices (a) and (b). In this case, the addition and the subtraction are partially reversed. (Not shown).

The above is a case in which the system is divided into a learning phase and a recognition phase. The speaker sequentially starts from speech uttered in the past (immediately) during a call or during recognition processing.
Communication or recognition can be performed while repeating learning. That is, (FIG. 5) to (FIG. 8), (FIG. 12) to
Buffer memory 5 in (FIG. 14), (FIG. 16), etc.
Reference numeral 1 denotes a state in which an input signal is always captured, and a motion vector is recalculated by the above-described method based on the voice data captured therein at appropriate intervals, thereby rewriting the codebook and normalizing the speaker. Can be updated. As a result, the speaker can perform speaker adaptation sequentially without special awareness of the learning phase, and can adapt or normalize following the temporal change of the speaker characteristic.

In the present embodiment, the movement vectors are calculated as h ₁ ,..., H ₂ ,..., H _M giving the extreme values of the objective function. I can do it. Further, in the present embodiment, an example of a case of obtaining the h _i to reduce the objective function, there is a case of obtaining the h _i to increase this by the definition of how the objective function. For example, if -J is used instead of J in the present example, this is naturally the case. In this embodiment, addition,
Although the word "subtraction" is used, if a negative sign is added, subtraction is addition and subtraction is addition, so any expression is valid on the word.

[0110]

As described above, according to the present invention, the codebook can be adapted to the speech of a specific speaker with a small number of samples, or the speech of the speaker can be adapted to the standard codebook. Normalization can be performed, and the communication quality can be improved in the case of communication and the recognition accuracy can be improved in the case of recognition with a small amount of learning effort.

[Brief description of the drawings]

FIG. 1 is a principle diagram of a transmission method based on vector quantization.

FIG. 2 is a general principle diagram of a speech recognition device based on vector quantization.

FIG. 3 is a detailed view of a reference model storage unit in FIG. 2;

Fig. 4 HMM (Hidden Markov Model) principle diagram

FIG. 5 is a principle diagram of an embodiment of an adaptation method according to the present invention.

FIG. 6 is a principle diagram of another embodiment according to the present invention.

FIG. 7 is a block diagram of a signal transmission device based on vector quantization based on the principle of FIG. 5;

8 is a block diagram of a signal transmission device based on vector quantization based on the principle of FIG. 6;

FIG. 9 is an embodiment of a receiving apparatus for the transmitting apparatus shown in FIGS. 7 and 8;

FIG. 10 is an embodiment of a receiving apparatus for the transmitting apparatus shown in FIGS. 7 and 8;

FIG. 11 shows another embodiment of the receiving apparatus for the transmitting apparatus shown in FIG.

FIG. 12 is a block diagram of a pattern recognition device based on vector quantization based on the principle of FIG. 5;

FIG. 13 is a block diagram of a pattern recognition device based on vector quantization based on the principle of FIG. 6;

FIG. 14 is a diagram of an embodiment of a transmission device based on speaker normalization.

FIG. 15 is a diagram of an embodiment of a receiving apparatus based on speaker normalization.

FIG. 16 is a diagram of an embodiment of a recognition device based on speaker normalization.

FIG. 17 is a diagram of another embodiment of the speaker normalization method according to the present invention.

[Explanation of symbols]

 1, 4 switch 2, 3 buffer memory 5 codebook 6 comparison section 7 label selection section 8 code vector reading section 9 codebook 52 movement vector calculation section 53 movement vector storage section

────────────────────────────────────────────────── ─── Continuation of the front page (56) References JP-A-3-75700 (JP, A) JP-A-4-122997 (JP, A) JP-A-4-127200 (JP, A) JP-A-5-127 173588 (JP, A) (58) Field surveyed (Int. Cl. ⁷ , DB name) G10L 15/00-19/14

Claims (24)

(57) [Claims] 1. Clustering a feature vector space ,
The representative vectors definitive in each cluster, and the standard code book for storing a searchable form with labels corresponding to each of the representative vectors of said standard codebook Modifa
To input the standard codebook into vector quantization.
And learning vector storage means for storing a number of learning vector for adapting the force vector, the standard code
By adding to the representative vector of the book,
The movement vector for modifying the table vector is
The representative vector of the quasi-codebook and the learning vector
Gives the extrema of the objective function defined as a function
A motion vector calculation means for calculating Te, and a codebook adaptation means to obtain a New representative vector by adding the motion vector in the representative vector, the input vector by the representative vector wherein A New are in encoding the input vector each label of the label or the input vector
A vector quantization device for converting a degree of belonging to a (cluster) into a degree of belonging vector having elements as elements. 2. Clustering a feature vector space ,
The representative vectors definitive in each cluster, and the standard code book for storing a searchable form with labels corresponding to each of the first input vector to be vector quantized Modi
File into the standard codebook.
To normalize the force vector to get a second input vector
And learning vector storage means for storing a number of learning vector of the adding to the first input vector
Therefore, the movement for modifying the first input vector
Let the vector be the same as the representative vector in the standard codebook
Objective function pole defined as a function of the learning vector
Movement vector calculation means for calculating a value to be given
And the motion vector is added to the first input vector.
Normalizing means for obtaining a second input vector by
When encoding the first input vector, the second input vector
Label or label the first input vector with a vector
Is the degree of membership of the input vector for each label.
A vector quantization device for converting the data into a membership vector . 3. The movement vector calculation means includes a standard codebook.
3. The vector quantization apparatus according to claim 1, wherein a motion vector is obtained for each of the representative vectors of the blocks . 4. The moving vector calculating means according to claim 1, wherein:
Tsu claim and obtains the moving vector as common to all the representative vectors of click 1 or claim 2
A vector quantizer as described. 5. The vector quantization apparatus according to claim 1, further comprising an adder for adding the movement vector and each representative vector of the standard codebook, and performing vector quantization by an output of the adder. 6. An adder for adding a motion vector to each representative vector of a standard codebook, and an adaptation codebook for storing an output of the adder, wherein an output of the adaptation codebook is used to generate a vector quantum vector. The vector quantization apparatus according to claim 1, wherein the vector quantization is performed. 7. A signal transmitting apparatus comprising: a label transmitting means for transmitting a label encoded by the vector quantization apparatus according to claim 5; and a moving vector transmitting means. 8. A label transmitting means for transmitting a label coded by the vector quantization apparatus according to claim 6, and an adaptive codebook transmitting means for transmitting an adaptive codebook. Signal transmission device. 9. A motion vector storage unit for storing a motion vector transmitted from the signal transmission device according to claim 7,
An adder for adding a standard codebook, a representative vector read from the standard codebook corresponding to the received label, and a movement vector read from the movement vector storage unit corresponding to the label; And a decoder that uses the output of the adder as a decoded vector of the label. 10. A motion vector storage unit for storing a motion vector transmitted from the signal transmission device according to claim 7, a standard codebook, a representative vector corresponding to each label of the standard codebook, and a corresponding vector. An adaptation codebook that stores the sum of each of the movement vectors read from the movement vector storage unit in association with the label, and the code vector of the adaptation codebook corresponding to the received label is stored as the code vector of the label. A signal receiving device including: a decoder that sets a decoded vector. 11. An adaptive codebook storage unit for storing an adaptive codebook sent from the signal transmitting apparatus according to claim 8, wherein a code vector of the adaptive codebook corresponding to a received label is stored in the adaptive codebook storage unit. A signal receiving device comprising: a decoding unit that decodes a label; 12. The vector quantization apparatus according to claim 1, wherein each vector of the input feature vector sequence is vector-quantized and converted into a label, and as a result, the feature vector sequence is converted into a label sequence. HMM storage means for storing a hidden Markov model (HMM) for each recognition unit in which the occurrence probability of each label is defined for each state, likelihood calculation means for calculating the likelihood of each HMM for the label sequence, A recognition apparatus, wherein a recognition unit corresponding to an HMM that gives the maximum likelihood is a recognition result. 13. An individual vector of an input feature vector sequence is vector-quantized and converted into a membership vector having the membership of each label as an element. As a result, the feature vector sequence is converted into the membership vector sequence. 2. A vector quantization device according to claim 1, wherein the probability of occurrence of each label is defined for each state, HMM storage means for storing a hidden Markov model (HMM) for each recognition unit, and said membership degree vector sequence. A likelihood calculating means for calculating the likelihood of each of the HMMs, and a recognition unit corresponding to the HMM giving the maximum value of the likelihood as a recognition result. 14. The vector quantization apparatus according to claim 1, wherein each vector of the input feature vector sequence is vector-quantized and converted into a label, and as a result, the feature vector sequence is converted into a label sequence. Recognition model storage means for storing a recognition model for each recognition unit expressed by a label sequence, distance calculation means for calculating the distance or similarity between the input label sequence and each recognition model, and a minimum value of the distance Alternatively, a recognition unit that uses a recognition unit corresponding to a recognition model that gives a maximum value of similarity as a recognition result. 15. An individual vector of the input feature vector sequence is vector-quantized and converted into a membership vector having the membership of each label as an element. As a result, the feature vector sequence is converted into a membership vector sequence. 2. The vector quantization device according to claim 1, wherein the recognition model storage unit stores a recognition model for each recognition unit expressed by a label sequence, and a distance or a similarity of each recognition model with respect to the input membership sequence. a distance calculation means for calculating a recognition device which is characterized in that the recognition result recognition unit corresponding to the recognized model giving the maximum value of the minimum or the similarity of the distance. 16. The vector quantization apparatus according to claim 2, further comprising an adder for adding the motion vector and the input vector, wherein an output of the adder is vector-quantized. 17. A signal transmitting apparatus comprising: a label transmitting means for transmitting a label encoded by the vector quantization apparatus according to claim 16, and a moving vector transmitting means for transmitting a moving vector. 18. A motion vector storage unit for storing a motion vector transmitted from the signal transmission device according to claim 17, a standard codebook, and a standard codebook read from the standard codebook corresponding to a received label. A subtractor for subtracting the motion vector read from the motion vector storage unit from the representative vector, and a decoder using the output of the subtractor as a decoded vector of the label. 19. The vector quantizer according to claim 2, wherein the normalized vector of each of the input feature vector sequences is vector-quantized and converted into a label, and consequently the feature vector sequence is converted into a label sequence. , An HMM storage unit that stores a hidden Markov model (HMM) for each recognition unit in which the occurrence probability of each label is defined for each state, and calculates the likelihood of each of the HMMs with respect to the membership vector series. A recognition apparatus characterized in that a likelihood calculating means and a recognition unit corresponding to an HMM giving the maximum value of the likelihood as a recognition result. 20. Vector quantization of a normalized vector of an individual vector of an input feature vector sequence, conversion into a membership vector having the membership of each label as an element, and as a result, A vector quantization device according to claim 2, wherein the vector quantization device converts the vector quantization sequence into a degree vector sequence;
HMM storage means for storing a hidden Markov model (HMM) for each recognition unit in which the occurrence probability of each label is defined for each state, and each HMM for the membership degree vector series
And a recognition unit that calculates a recognition unit corresponding to the HMM that gives the maximum value of the likelihood as a recognition result. 21. The vector quantizer according to claim 2, wherein the normalized vector of each of the input feature vector sequences is vector-quantized and converted into a label, and consequently the feature vector sequence is converted into a label sequence. And a recognition model storage means for storing a recognition model for each recognition unit expressed by a label sequence, and a distance calculation means for calculating the distance or similarity between a pair of the input label sequence and each of the recognition models. A recognition unit corresponding to a recognition model that gives the minimum value of the distance or the maximum value of the similarity as a recognition result. 22. Vector quantization of a normalized vector of an individual vector of an input feature vector sequence is performed, and the normalized vector is converted into a membership vector having the membership of each label as an element. A vector quantization device according to claim 2, wherein the vector quantization device converts the vector quantization sequence into a degree vector sequence;
Recognition model storage means for storing a recognition model for each recognition unit represented by a label sequence, distance calculation means for calculating the distance or similarity of each recognition model with respect to the input membership degree series, and a minimum value of the distance Alternatively, a recognition apparatus characterized in that a recognition unit corresponding to a recognition model that gives the maximum value of similarity is a recognition result. 23. Temporary storage means for sequentially storing a predetermined constant signal section of an input signal, wherein the contents of said temporary storage means are used as a learning vector and a codebook or a moving picture is sequentially stored for each signal section. The vector quantization apparatus according to claim 1, wherein the vector quantization is performed. 24. Temporary storage means for sequentially storing a predetermined constant signal section of an input signal, and using the contents of said temporary storage means as a learning vector, the normalization of the input signal is sequentially performed for each signal section. The vector quantization apparatus according to claim 2, wherein a motion vector for quantization is calculated.