JP2545960B2 - Learning method for adaptive speech recognition - Google Patents

Description

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to learning of a speech recognition apparatus for recognizing speech, which is adaptable to different speakers and utterances in different utterance noise environments.

(Prior Art) Conventionally, there has been a method of recognizing an input voice by using a standard pattern that is held in advance. (See Kyoritsu Shuppan Co., Ltd., “Voice Recognition” (Reference 1) p.101-113) In this method, when a voice of a speaker other than the speaker used to create the standard pattern is recognized, the difference in the voice pattern of the speaker Therefore, a sufficient recognition rate cannot be obtained. In addition, even if the speaker is the same, if the utterance environment such as ambient noise is significantly different from that when the standard pattern is created, the recognition rate will decrease.

Usually, when recognizing the voice of a specific speaker, it is necessary to utter all the recognition target words and register them. However, it has a drawback that it requires a lot of labor when the vocabulary is large. On the other hand, a method of adapting a standard pattern to a specific speaker with a small amount of voice has been proposed. As an example, IEEE
ICASSP-86,49.5p.2643 “Speaker Adaptation through V
The speaker adaptation method by vector quantization is described in "ector Quantization" (Reference 2). In addition, as a method for improving the performance deterioration due to the influence of the quantization error due to the vector quantization in this method, a patent application is proposed. A speaker adaptation method using a neural network is described in Sho 63-122559. In these methods, a pattern in which speaker 1 and speaker 2 utter the same word (a phrase or a sentence may be used) is DP. The matching is performed for optimum temporal correspondence, and then the adaptation, that is, the converter is obtained from the set of corresponding feature vector pairs.

(Problems to be Solved by the Invention) In the above method, voice patterns of different speakers are matched as they are, but the time axis correspondence by DP matching is not always accurate. For example, if one element of speaker 1 resembles another phoneme of speaker 2 or the like, incorrect correspondence may occur. Such an error reduces the performance of adaptation and the recognition performance. An object of the present invention is to eliminate such errors in time axis correspondence due to speaker differences and realize highly accurate speaker adaptation and vocal noise adaptation.

(Means for Solving the Problem) In the learning method for adaptive speech recognition according to the present invention, a standard pattern of environment 1 is converted into an environment by a neural network for environment adaptation learned from the same utterance pattern of environment 1 and environment 2. Two
In the method of recognizing using the pattern converted for
By repeating the process of correcting the weighting factor of the neural network by using the error pattern between patterns obtained by associating the pattern A obtained by converting the learning pattern of environment 1 with the neural network and the learning pattern B of environment 2 by the optimum time axis correspondence. It is characterized by having a means for learning a neural network.

(Operation) The operation of the present invention will be described by taking speaker adaptation as an example. It is assumed that there is a pattern A of two speakers 1 and a pattern B of a speaker 2 having the same utterance content, and a neural network for converting pattern A to pattern B is learned. The patterns A and B are represented by a vector time series A = {a (i), i = 1, I} B = {b (j), j = 1, J}, and the iteration step of learning is represented by k. The neural network realizes conversion between the vector of pattern A and the vector of pattern B, and both input and output are vectors. The vector at each time of the pattern A is converted by the neural network of step _k to obtain the pattern B ^* _k .

B ^* _k = {b ^* _k (i), i = 1, ..., I} DP between input pattern conversion pattern B ^* _k and teacher pattern B
Match. DP matching solves the following minimization problem. Details of DP matching are described in Reference 1.

At this time, the optimum J (i) is also obtained. The pattern B ^* pattern matched to the time axis of _k B to B _k.

_{B k = {b k (i} ), i = 1, ..., I} the d _k error function time series pattern of the error vector between the B _k of the pattern B ^* _k and D _k.

Backpropagation learning is performed using the error vector time series pattern d _k to correct the weight of the neural network. The details of backpropagation learning are described in the Institute of Electronics, Information and Communication Engineers, "Speech by Stochastic Model," p.164-167.

Although the correction will be performed using I error vectors, as a correction method, a method of simply repeating the correction with the load correction amount obtained for one error vector,
Alternatively, as described in Reference 2, a method of calculating the load correction amount for I error vectors and then averaging them to correct the load is used. In this way, the neural network of step k + 1 is obtained. Due to the convergence of the back propagation learning, the error function of the neural network at step k + 1 is smaller than the error function of the neural network at step k + 1 under the condition fixed to the teacher pattern B _k .

That is, Holds. On the other hand, since the DP matching executes the matching that minimizes the error function among all possible time axis matching, the error function D _{k + 1} obtained as a result of the DP matching in step k + 1 is Holds. Since D _{k + 1} ≦ D _k is established from the equations (1) and (2), the above iterative processing converges, and the neural network can be learned by the above iterative processing.

It is possible to give a random value as the initial value of the weighting factor of the neural network in learning, but the pattern A and B are directly DP matched to the time axis by the method described in Japanese Patent Application No. 63-122559. From the result of attaching
It is also possible to perform learning of the neural network and use the obtained result as the initial value.

As the patterns A and B, words, sentences, or a plurality of word sets, sentence sets, or the like having the same utterance content can be used.

Further, by using the utterance patterns of the same speaker having different environmental noises as the patterns A and B, it is possible to adapt the utterance environmental noises.

(Example) FIG. 1 is a diagram showing an example according to the present invention. In the figure, storage units 1 and 2 hold patterns A and B, respectively, and a learning control unit 3 generates a control signal k representing a learning step. First, the pattern A held in the storage unit 1 is input to the neural network unit 4 and converted into the pattern B ^* _k . Then, the pattern B held in the storage unit 2 and the pattern B ^* _k are D
It is input to the P matching unit 5. The DP matching unit 5
Two patterns B input, executes the DP matching between the B ^* _k, and outputs the pattern B _k which are matched pattern B to pattern B ^* _k. The error pattern calculation unit 6 calculates an error pattern d _k between the pattern B _k output from the DP matching unit 5 and the pattern B ^* _k output from the neural network unit 4. The error pattern d _k is sent to the neural network correction unit 7 and the error function calculation unit 8. The neural network correction unit 7 corrects the content (weighting factor) of the neural network unit 4 by back propagation learning. The learning control unit 3 performs control to repeat the above series of operations until the error function D _k calculated by the error function calculation unit 8 becomes a certain value or less or step k becomes a predetermined value or more. .

(Effects of the Invention) According to the present invention, it is possible to learn a neural network that can effectively adapt to a new speaker or a vocal noise environment,
A high-performance and adaptive voice recognition device can be realized.

[Brief description of drawings]

FIG. 1 is a diagram showing an embodiment according to the present invention, in which 1 and 2 are a pattern storage unit, 3 is a learning control unit, 4 is a neural network unit, 5 is a DP matching unit, and 6 is an error pattern calculation. Reference numeral 7 is a neural network correction unit, and 8 is an error function calculation unit.

Claims (1)

(57) [Claims] 1. An environment 1 is created by a neural network for environment adaptation that is learned from the same speech patterns of environment 1 and environment 2.
In the speech recognition learning method for recognizing using the pattern obtained by converting the standard pattern of the environment 2 into the environment 2, the learning pattern of the environment 1 is converted into the pattern A and the environment 2 by the neural network.
The adaptive speech having means for learning the neural network by repeating the process of correcting the weighting factor of the neural network using the error pattern between the patterns obtained by associating the learning pattern B with the optimal time axis. A learning method for recognition.