JP2996417B2 - Voice recognition 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 method for recognizing speech by using a so-called feed-forward artificial neural circuit in which an input is propagated in one direction to obtain an output, that is, a so-called neural network.

[0002]

2. Description of the Related Art Among artificial neural circuits, a multilayer perceptron type uses a reverse error propagation method, so-called back propagation, by introducing a differentiable nonlinear function, so-called sigmoid, as an output function of a cell corresponding to a neural element. High-precision learning that has been used [DE Rumelhart, et al, “Learning Internal Repr.
esentations by Error Propagation ", Parallel Distri
buted Processing: Explorations in the Microstructu
re of Cognition. Vol. 1: Foundations. MIT Press (198
6)]. The multilayer perceptron type neural network is called a feedforward type neural network and has been applied to speech recognition. Conventionally, a time delay neural network (TDN) has been proposed as a method for constructing a neural network that is resistant to temporal displacement of speech features.
N) has been proposed [AHWaibel, et. Al, “Phone
me Recognition Using Time-Delay Neural Network, "I
EEE Trans., ASSP Vol. 37, No.3, pp.328-339, (Mar,
1989)]. The feature of the time-delay neural network is that the connections in the neural network arranged in the time direction are tied, that is, the same coupling coefficient. However, in conventional feedforward neural networks for speech recognition, including time-delay neural networks, the input frame length is fixed, and it expands and contracts in time compared to the speech used for learning, regardless of linear non-linearity. The recognition rate for the unknown input speech was low. In order to partially absorb time expansion and contraction, a time-structured neural network that divides the tied connection of a time-delay neural network into several parts in time has been proposed. [Komori et al., “Phonological Recognition Using Neural Networks Considering Time Structure,” Proc. Of the Acoustical Society of Japan Spring Meeting, Vol.1, pp.157-158. (Ma
r, 1990)].

[0003]

[Problems to be solved by the invention] Speech is personality, context,
Depending on the utterance speed, local expansion and contraction of the time axis, that is, non-linear expansion and contraction is caused. Using a feed-forward type neural network with a fixed input data length but excellent pattern discrimination performance, the present invention realizes high recognition by recognizing data of various lengths in consideration of the non-linear expansion and contraction of the time axis. The goal is to achieve performance.

[0004]

According to the present invention, for a given voice section, a plurality of sets of feature parameter time series are generated in accordance with a plurality of predetermined time expansion / contraction functions, and these are input to a neural network. From the corresponding time point in the set of these feature parameter time series, a set of connections to the neural cell of the first hidden layer is a so-called tied connection in which constraints are set so as to have the same connection coefficient. It is characterized in that a higher-order cell is fired if it matches any of the expansion / contraction patterns of various time expansion / contraction, and that the neural network has a structure that can accept various time-expanded voices. This structure of the neural network will be referred to as a time-reducing neural network.

[0005]

According to the method of the present invention, a time-expandable neural network has a structure capable of receiving a plurality of types of time-expanded voices. When an unknown voice is input to the neural network, a cell of the first hidden layer corresponding to any one or more of a plurality of time expansion / contraction patterns is fired,
Since they are integrated in the upper layer, it is possible to recognize time-expanded speech.

[0006]

An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 shows an example of a phoneme recognition system to which an embodiment of the present invention can be applied. In using this system, first, both switches SW1 and SW2 are set to b
Flip to the side to learn the neural network. A standard voice for neural network learning is input from the microphone 1,
The output of the microphone 1 is passed through a filter 2 having a band half the sampling frequency at the time of A / D conversion, and the output of the filter 2 is converted into a digital value by an A / D converter 3. In this embodiment, the sampling frequency is 12 kHz, but the sampling frequency may be different. Next, the output of the A / D converter 3 is passed through a mel-scale bandpass filter 4 to obtain a plurality of characteristic time series. In this embodiment, a mel-scale band-pass filter group of 16 channels is used, but the number of channels may be different.
In addition to the feature extraction by the mel-scale bandpass filter, a parameter representing a spectrum such as a cepstrum coefficient may be used. Various methods are conceivable for the design of the mel-scale bandpass filter. In this embodiment, the sum of the outputs of several channels is obtained based on the mel scale from the output of 128 channels obtained by a 256-point fast Fourier transform so-called FFT. And use the logarithm.

Next, the phoneme label corresponding to the standard speech stored in the phoneme label 6 is passed through the phoneme position determination unit 7 and supplied to the phoneme cutout unit 5, and the phoneme label is output from the time series of the output of the bandpass filter 4. Cut out the phoneme part based on. A predetermined time expansion and contraction is performed on the cut-out phoneme part by the time expansion and contraction unit 8 to generate a time-series set of feature parameters. In this embodiment, five types of time expansion / contraction functions are prepared. These functions y are represented by a time axis obtained by normalizing the time x from 0 to 1 and a phoneme section length obtained by the phoneme position determining unit 7 as T. y = Tx y = T (x ± 0.3 sin (πx)) y = T (x ± 0.15 sin (2πx)). These time expansion functions y are shown by lines 21 to 25 in FIG. Any other monotonically increasing function can be used as the time stretching function. The first equation (line 21) corresponds to a linear stretching function. As the characteristic parameter time series, the time points of y corresponding to three points where x is 1/6, 3/6, and 5/6 are obtained, and three frames before and after the three time points as the center, that is, nine frames in total are used. For example, looking at the line 21, there are 9 frames indicated by diagonal lines in FIG. 2, the line 25 is 9 frames of 3 frames each separated by a dotted line, and the line 25 is an example in which the time axis is compressed at the center. is there. The number of points and the number of frames can be increased or decreased. In this embodiment, since one frame includes the output of the band-pass filter 4 of 16 channels and the number of the time expansion / contraction functions is 5, the input number of the neural network is 16 × 9 ×
5 = 720. The network generation unit 11 creates a structure of a time-expanded neural network according to the number of input cells, the number of hidden cells, the number of output cells, the number of time expansion / contraction, the number of time axis sample points, and the like.

FIG. 2 shows the structure of a time-varying neural network constituting the phoneme recognition unit 9. In the input layer 26, the first hidden layer 27, the second hidden layer 28, and the output layer 29, nerve cells are arranged in a matrix. A vertical band called a column in a matrix is referred to as a frame here. The connection between the layers indicates that all the cells of the lower layer frame group and all the cells of the upper layer frame are all connected. The connection of the cell group and the cell group in all combinations is called full connection. For example, the connection on the left side between the input layer 26 and the first hidden layer 27 indicates that three frames of the input layer 26 and one frame of the first hidden layer 27 are fully connected. Figure 2
Indicate that the corresponding connections of the five full connections indicated by the thick arrows are forcibly learned to have the same coupling coefficient. That is, the connection from the corresponding position of the five time expansion / contraction patterns is tied, and these are multiplexed into twelve. In the second layer 28, these are multiplexed into 12 pieces. The outputs of all the cells of the second hidden layer 28 are integrated to fire the output cells. The input / output function of each cell is a standard sigmoid. That is, a certain cell j
An output p _i of cell i of the lower layer to be input to the coupling coefficient w _ji
Then, the output q _{j of the} cell j is obtained by q _j = 1 / [1 + exp (− (Σ _i w _ji p _i + bias))]. Σ _i indicates the sum for all cells i entering cell j, and bias is the coupling from the special input cell that supplies the DC bias. The bias to the tied cell is still tied. In this embodiment, a four-layer network is used. However, a three-layer network in which the output of the first hidden layer is fully connected to all output cells may be used.

The neural network learning unit 10 in FIG. 1 uses the phoneme data for learning obtained by the time expansion and contraction unit 8 to determine the network coupling coefficient of the phoneme recognition unit 9 by the backward error propagation method, so-called back propagation. . The learning of the tied combination is performed by updating the coupling coefficient by averaging the coupling coefficient correction amount in the tied coupling set.

After the learning for the phoneme recognizing section 9 is completed in this way, the switch SW1 is used to recognize the unknown phoneme.
And SW2 are switched to the a side to input a sound from the microphone 1, obtain the output of the mel-scale bandpass filter 4 in the same processing system as in the learning, and determine the phoneme position based on the inspection or the volume by the phoneme position determination unit 7. The phoneme is cut out from the determined position in the phoneme cutout unit 5. The time expansion and contraction of the phoneme is performed by the time expansion and contraction unit 8 to obtain a set of feature parameter time series, and its output is given as an input to the neural network of the phoneme recognition unit 9. It recognizes whether the cell has produced the largest output, that is, has fired. Recognition result display section 1
2 displays the recognition result.

In this embodiment, phonemes are recognized, but voices of any length, such as syllables and words, can be recognized with the same configuration. However, it is necessary to adjust the number of cells in each layer of the neural network to be used according to a set of time expansion / contraction functions and the length of input speech.

[0012]

According to the present invention, the effect of the present invention is expressed by six phonemes / b, d, g, m, n, N /
Was confirmed by a recognition experiment. The phonemes used for learning were extracted 200 times from the even-numbered 5240 frequently used important words. The phonemes used in the test were extracted by inspection from 115 sentences uttered by the same speaker as the learning speech in sections. As a method that is an extension of the conventional method, data obtained by linearly expanding and contracting speech in a given section and resampling the data into a fixed number of frames is converted to a time-delay neural network (TDN).
N), the phoneme recognition rate was 81.3%, but when the method of the present invention was used, the recognition rate was 8%.
It was improved to 4.5%.

As described above, according to the present invention, a high recognition rate can be obtained by inputting an input voice while expanding and contracting it in time with a plurality of expansion and contraction patterns.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an example of a phoneme recognition system to which an embodiment of the present invention has been applied.

FIG. 2 is a block diagram showing a configuration example of a time-varying neural network which is a main part of the present invention.

Continuation of the front page (56) References JP-A-2-77888 (JP, A) JP-A-1-241668 (JP, A) JP-A-58-115487 (JP, A) Spring Study of the Acoustical Society of Japan in 1991 Proceedings of the conference ▲ I ▼ 1-5-8 “Consonant recognition using time-expandable neural networks” p. 19-20 (issued March 27, 1991) IEICE Technical Report [Voice] Vol. 91, No. 95, SP91-13, “Speech Recognition Using Time-Expandable Neural Network” p. 55-62 (Issued June 20, 1991) (58) Fields investigated (Int. Cl. ⁶ , DB name) G10L 3/00 531 G10L 3/00 539 JICST file (JOIS)

Claims (1)

(57) [Claims] 1. A method for performing speech recognition using a feedforward neural network in which an input is propagated in one direction and an output is obtained, wherein a method is provided in which a given speech is expanded and contracted on a plurality of time axes. A set of feature parameter time series at a certain number of time points is input, and the first time from the feature parameter at the corresponding time point in the set of feature parameter time series
A set of connections to the neural cells of the hidden layer is a tied connection with constraints so as to have the same connection coefficient, and a plurality of cells of the first hidden layer to be connected from the feature parameters at one time point of the input are prepared. A speech recognition method characterized by performing recognition based on an output of an output cell when a set of feature parameter time series obtained by performing the above-described time expansion and contraction on speech is input to the neural network.