JPH0752356B2 - Speaker adaptation method - Google Patents

Description

[Detailed description of the invention]

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speaker adaptation method, and more particularly, in the field of speech recognition, using training samples generated by an unknown speaker, a feature vector of a standard speaker and a feature vector of an unknown speaker are obtained. It relates to a speaker adaptation method that obtains a correspondence relationship between and performs adaptation based on the obtained correspondence relationship.

[0002]

2. Description of the Related Art Conventionally, the speech of a standard speaker and the speech of an unknown speaker are matched on the time axis by DP matching, and the correspondence relationship between the feature vector of the standard speaker and the feature vector of the unknown speaker is found in speaker adaptation. In the adaptation method, the correspondence obtained from the learning data is used as it is for adaptation.

[0003]

However, if the amount of learning data is small, the correspondence obtained therefrom will have variations depending on the data, and the reliability of the correspondence will be low. That is, when the amount of learning data is small, there is a problem that the learning data is biased toward the small amount of learning data, and the correspondence relationship depends on the biased data.

[0004] Therefore, the main object of the present invention is to perform smoothing in consideration of the continuity of the speaker's space in order to absorb variations contained in correspondences obtained from learning, thereby obtaining more accurate correspondences. It is to provide such a speaker adaptation scheme that can be obtained.

[0005]

[Means for Solving the Problems] The invention according to claim 1 is
The speech of a standard speaker is vector-quantized and stored as a codebook, and in response to input of the feature vector of an unknown speaker, the correspondence relationship between the feature vector of the unknown speaker and the feature vector of the standard speaker. are matched using the codebook, and the codebook is updated by calculating the average of the feature vectors of the input speaker matched with the feature vectors of the standard speaker. Based on a fuzzy function, feature vectors are smoothed and stored in a codebook.

[0006]

[0007]

[0008]

[0009]

In the speaker adaptation system according to the present invention, by smoothing the correspondence obtained from the learning samples, even if only a small amount of data for learning is obtained, it is possible to obtain the correspondence with high accuracy. It is possible to achieve highly accurate speaker adaptation.

[0010]

DETAILED DESCRIPTION OF THE INVENTION FIG. 1 is a schematic block diagram of a speech recognition apparatus to which the present invention is applied. In FIG. 1, the speech recognition apparatus includes an amplifier 1, a low-pass filter 2, and an A/D converter 3.
and a processing device 4 . An amplifier 1 amplifies an input speech signal, and a low-pass filter 2 removes aliasing noise from the amplified speech signal. The A/D converter 3 converts the audio signal into a 16-bit digital signal using a 12 kHz sampling signal. The processing device 4 includes a computer 5 , a magnetic disk 6 , terminals 7 and a printer 8 . Based on the voice digital signal input from the A/D converter 3, the computer 5 performs adaptation between speaker feature vectors using the method shown in FIG. 2, which will be described later.

FIG. 2 is a flowchart for explaining the operation of one embodiment of the invention. Standard speaker codebook C ^R
is stored on the magnetic disk 6 shown in FIG. 1, and this standard speaker's codebook C ^R is used as the initial value of the transformation codebook C ^T . Step (abbreviated as SP in the figure) S
At P1, when the input vector sequence of the unknown speaker is input, at step SP2, the input vector sequence of the unknown speaker and the code sequence of the standard speaker are converted into a DTW (dynamic time warping method) using a conversion codebook C ^T . ) is used to make the correspondence.

At step SP3, step SP2
Update the transform codebook C ^T from the correspondence obtained in . That is, the transformed codebook C ^T is obtained by calculating the average of the input speaker's vectors associated with the standard speaker's code vectors. At that time, using the sub-vectors of the code vectors for which the correspondence has already been determined,
An estimation of the difference vector of the unmatched vectors is performed. More specifically, let M be the set of input vectors associated with the m-th vector C ^T _m , and let the difference vector V _m between the average value of the input vectors x belonging to the set M and the vector C ^R _m be: Calculated by Equation 1.

[0013]

[Number 1]

N._n= 0 vector C^R. _nabout,
N._kA code vector C with >0^R. _kfuzzy class function of
μ_n,kAsk for Vector C^R. _ndifference vector V_nof
Vector C^R. _kdifference vector V_kand μ_n,kUsing,
It is calculated from the following formula 2.

[0015]

[Number 2]

[0016] All vectors C ^T _n in the transform codebook C ^T are updated using the difference vector V according to Equation 3 below.

[0017]

[Number 3]

The correspondence obtained as described above is D
It is obtained from a small number of words within the framework of TW, and does not necessarily represent the correspondence in the different speaker space itself, and contains errors. On the other hand, considering the continuity of the speaker space, it is natural to think that the correspondence in the different speaker space is also continuous. Therefore, in step SP6,
In order to absorb the error contained in the DTW correspondence and obtain the true correspondence, smoothing based on the fuzzy series function is performed as follows. That is, in step SP5, the fuzzy series function μ _n,k : _k ≠ _n of all vectors C ^R _k for which N _k >0 for vectors C ^R _n is obtained. Next, the difference vector of the code vector C ^R _n is calculated by the following equation (4).

[0019]

[Formula 4]

[0020] Here, N _k is considered as the reliability of the correspondence and is used as a weight for the difference vector. α is a constant for making the contribution of N _k the same as μ _n,k , and β is the predetermined reliability of V _n . Update all transform codebooks with the smoothed difference vector _V'n .

By varying the fuzziness during smoothing, the space in which continuity is taken into account can be controlled. That is, the closer the fuzziness is to 1, the more local the space is.
The closer to ∞, the more global space is considered. Therefore, when there is a large amount of training data and the matching is sufficiently reliable, a fuzzyness close to 1 is used, and when there is only a small amount of training data and the matching is unreliable, a large fuzzyness is used to achieve higher accuracy. high speaker adaptation can be realized.

[0022]

INDUSTRIAL APPLICABILITY As described above, according to the present invention, by performing smoothing while paying attention to the continuity of the speaker's space, variations in correspondences obtained from learning can be absorbed, resulting in more accurate correspondences. A relationship can be obtained, and highly accurate speaker adaptation can be achieved even when only a small amount of training data is obtained.

[Brief description of the drawing]

1 is a schematic block diagram of one embodiment of the present invention; FIG.

FIG. 2 is a flowchart for explaining specific operations of one embodiment of the present invention;

[Description of symbols]

1 amp 2 Low pass filter 3 A/D converter 4 processing equipment 5 computer 6 magnetic disk 7 Terminals 8 Printer

Claims (1)

[Claims] [Claim 1] A speech of a standard speaker is vector-quantized and stored as a codebook, and in response to input of a feature vector of an unknown speaker, the feature vector of the unknown speaker and the standard speaker are generated. using the codebook, updating the codebook by obtaining an average of the feature vectors of the input speaker associated with the feature vectors of the standard speaker; A speaker adaptation method characterized in that, if an error is included in the matching, the feature vectors are smoothed based on a fuzzy function and stored in the codebook.