JPH04369699A - Unspecified spealer voice recognition system using neural network - Google Patents

Abstract

PURPOSE:To offer the neural network which extending a neural network architecture, proposed for the recognition of a specific person or a limited number of speakers, so that the voice of the specific speaker can be recognized. CONSTITUTION:A voice feature quantity is inputted to an input layer 3 in the form of a feature parameter time sequence, and propagated to 1st hidden layers 40, 41...4n in parallel at the same time to extract features characteristic to respective speakers. The 1st hidden layer 40 extracts effective feature quantities for discriminating between the speakers at the same time. The extracted features are propagated to 2nd hidden layers 50, 51, 52...5n, the network is learnt by an error reverse propagating method, and the result is outputted to an output layer 6.

Description

[Detailed description of the invention]

0001

[Industrial Application Field] The present invention relates to a method for recognizing speaker-independent speech using a neural network, and in particular, to a neural network applied to the field of speech recognition technology that uses a neural network to perform speech recognition for an independent speaker. This paper relates to a speaker-independent speech recognition method.

0002

BACKGROUND OF THE INVENTION In recent years, neural networks have been actively applied in the field of speech recognition. In particular, voiced plosives /b, d, g/
Since its high performance in phoneme recognition was demonstrated, TDNN
A network for recognizing 18 consonants with the basic structure and 23
Many networks for phoneme recognition and networks for multi-speaker phoneme recognition have been proposed.

[0003] However, a system that can fully recognize the speech of unspecified speakers from the level of phoneme recognition has not yet appeared. However, those that perform phoneme recognition for a limited number of speakers, such as Hampshire
J. , and A. Waibel: “The
Meta-Pi Network: Connecti
onist Rapid Adaptation for
r High Performance Multi-
SpeakerPhoneme Recognition
in ”, Proceedings of the
1990 IEEE International C
onference on acoustics, S
peach and signal process
ng, S3.9, pp164-168, 1990
．． It has been proposed in However, the performance of these recognition systems for the voices of unknown speakers, who are different from the learned speakers, has not been verified.

[0004] Therefore, the main object of the present invention is to provide a speaker-independent speech recognition method using a neural network that can reduce the learning time and the number of samples and can perform highly accurate recognition.

[0005]

[Means for Solving the Problems] The present invention configures a single network by integrating a network learned between each speaker and a network for speaker identification learned to discriminate between speakers. , the entire network is constructed by additional learning.

[0006]

[Operation] The speaker-independent speech recognition method using a neural network according to the present invention integrates a network trained for each speaker and a network for speaker identification trained to distinguish between speakers. By training each network individually, the training time and number of samples can be reduced, making highly accurate recognition possible.

[0007]

DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 is a schematic block diagram of an embodiment of the present invention. Referring to FIG. 1, a voice input signal is given to a feature analysis section 1, where FFT analysis and LPC analysis are performed, and then given to a neural network 2, which is a feature of the present invention, where voice recognition is performed and a recognition result is obtained. Output.

FIG. 2 is a concrete block diagram of the neural network shown in FIG. 1. Referring to FIG. 2, the neural network has an input layer 3 and a first hidden layer 40, 4.
1, 42...4n, and the second hidden layer 50, 51, 52...5
n and an output layer 6. The first hidden layer 41 is a sub-network that is trained using training samples of speaker 1, and the second hidden layer 51 is a sub-network that is trained using training samples of the same speaker 1. 42 is a sub-network that is trained using the learning samples of speaker 2, and the second hidden layer 52 is a sub-network that is trained using the learning samples of the same speaker 2. The first hidden layer 4n is the speaker N.
is a subnetwork trained with training samples of
The second hidden layer 5n is a subnetwork of the same speaker N. The first hidden layer 40 is a sub-network called a speaker identification network that uses learning samples from speakers 1 to N to determine which speaker's phoneme it belongs to. The output layer 6 finally determines the phoneme categories C1, C2,...Ck...CK from the values of each output unit.

Next, the operation of one embodiment of the present invention will be explained. The audio features input in the input layer 3 in the form of feature parameter time series are input to the input layer 3 and the first hidden layer 41, 4.
2...4n in parallel and simultaneously to the first hidden layers 41, 42...4n. At this time, each subnetwork extracts features specific to each speaker for each subnetwork of each speaker, and at the same time,
The first hidden layer 40 simultaneously extracts feature amounts effective for distinguishing between speakers.

Next, the first hidden layer 40, 41, 42...4
The output of n is transmitted to the second hidden layer 50, 51, . 52...
5n. Hidden layer second layer 50, 51, 52...5
The connection from n to the output layer 6 is such that the kth sublayer of each speaker's subnetwork is connected to the unit corresponding to the kth category Ck of the output layer 6, as shown in FIG. The speaker identification network is also connected in the same way, but the second hidden layer 50, 5
The connections from 1, 52, . . . , 5n to the output layer 6 are full connections. Furthermore, in order to maintain modularity, each subnetwork is not connected. This network is trained using the error backpropagation method (McClel
land J. L. , D. E. Rumelhar
t and the PDP Research Group
oup: “Parallel Distribution
d Processing”, vol.1. Ch.
ap. 8. MITPress (1988). ) can be done.

[0011] A network that integrates a network learned for each speaker and a network for speaker identification as described above has a high modularity, so that learning can be performed for each sub-network, which has been conventionally Compared to the proposed network and a simple four-layer network with the same degree of freedom (number of network connections), this has the advantage of significantly reducing learning time and training samples. In addition, the recognition rate is also stable and high. Satoru Nakamura, Hidefumi Sawai: "Study of neural network architecture for speaker-independent phoneme recognition" Institute of Electronics, Information and Communication Engineers Speech Research Group, SP90-61, 1990
Experimentally proven on December 20th.

0012

[Effects of the Invention] As described above, according to the present invention, the configuration of the neural network is configured into a module consisting of a subnetwork for each speaker and a network for speaker identification, and the learning of each subnetwork is performed individually. This makes it possible to reduce learning time and the number of samples, making highly accurate recognition possible.

[Brief explanation of the drawing]

FIG. 1 is a schematic block diagram of an embodiment of the invention.

FIG. 2 is a specific block diagram of the neural network shown in FIG. 1.

[Explanation of symbols]

1 Feature analysis unit 2 Neural network 3 Input layer 40, 41, 42...4n Hidden layer 1st layer 50, 51,
52...5n Hidden layer 2nd layer 6 Output layer

Claims (1)

[Claims] Claim 1: A network learned between each speaker and a network for speaker identification learned to discriminate between speakers are integrated to form a single network, and the entire network is constructed by additional learning. A speaker-independent speech recognition method using a neural network, which is characterized by comprising: