JPH05119798A - Word recognition system - Google Patents

Abstract

PURPOSE:To provide the word recognition system which is tolerant to the nonlinear expansion or contraction of a speech in the time direction due to difference in voicing custom among speakers and variance in speaker's voicing and both reduced in arithmetic quantity (speeded up) and improved in recognition precision. CONSTITUTION:The word recognition system 10 which recognizes words from an input speech by using a neural network 17 finds power from the input speech, frame by frame and then frames corresponding to plural peaks from a time series of the power, discriminates between the found frames and other frames and divides the speech into blocks of frames between the start point and peak of the speech, the peak and peak, and the peak and end point, and inputs the mean frequency characteristic of the speech in each block to the neural network 17.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a word recognition method using an average of frequency characteristics corresponding to input phonemes as an input for recognition and a neural network as a recognition method.

[0002]

2. Description of the Related Art In a conventional word recognition method using a neural network, as described in Japanese Patent Application No. 1-98376, a characteristic parameter of an input voice is calculated for each frame and a voice segment equally divided in time is obtained. As one block, the average of characteristic parameters in the block is used as an input parameter.

[0003]

In the conventional method, the input speech is equally divided into blocks without considering the nonlinear expansion and contraction of the speech and phonological divisions, and the average of the characteristic parameters in the blocks is used as the input parameter. Therefore, it is difficult to reduce the amount of calculation (speeding up) and it is difficult to recognize similar words because it is vulnerable to the difference in the utterance habits between the speakers and the variation in the utterance within the speaker.

The present invention is resistant to non-linear expansion and contraction of speech in the time direction due to differences in vocalization habits among speakers and variations in vocalizations within speakers, and at the same time reduces the amount of calculation (speeding up) and improves recognition accuracy. An object is to provide a word recognition method that can be achieved.

[0005]

According to a first aspect of the present invention, in a word recognition system for recognizing a word from an input voice by using a neural network, the power is calculated for each frame from the input voice, and when the power is the power. The frame corresponding to a plurality of peaks is obtained from the sequence, the obtained frame and the other frames are distinguished, and the block of the frame corresponding to the start point and the peak of the voice, the block of the frame corresponding to the peak, and the peak. The value obtained by dividing the voice into the blocks of the frame corresponding to the peak and the blocks of the frame corresponding to the peak and the end point and calculating the average of the frequency characteristics of the voice in each block It is intended to be input to the network.

According to a second aspect of the present invention, in addition to the first aspect of the present invention, a neural network having an input unit corresponding to the number of divisions of the input voice is used as the neural network. is there.

According to a third aspect of the present invention, in addition to the first aspect of the present invention, one neural network is used as the neural network.

[0008]

The redundant features (frequency characteristics) are averaged and replaced by one representative point, and the input voice is divided in consideration of the simple phoneme classification. It is resistant to non-linear expansion and contraction of speech in the time direction due to variations in utterances, and it is possible to reduce the calculation amount (speed) and improve recognition accuracy at the same time.

However, the description of the "neural network" in the present invention is as follows (1) to (4).

(1) From the structure of the neural network, the neural network and the hierarchical network shown in FIG.
It can be roughly classified into two types of mutual connection networks shown in (B). The present invention may be configured by using either of both networks, but the hierarchical network is more useful because a simple learning algorithm as described later has been established.

(2) Network Structure As shown in FIG. 5, the hierarchical network has a hierarchical structure including an input layer, an intermediate layer, and an output layer. Each layer is composed of one or more units. The coupling is only forward coupling such as input layer → middle layer → output layer, and there is no coupling in each layer.

(3) Structure of the unit The unit is a model of a brain neuron as shown in FIG. 6, and the structure is simple. Receive input from other units,
The sum is taken and converted according to a certain rule (conversion function), and the result is output. A variable weight, which represents the strength of the bond, is attached to each of the bonds with other units.

(4) Learning (Back Propagation) Learning a network is to bring an actual output close to a target value (desired output). Generally, the conversion function and weight of each unit shown in FIG. Is learned by changing.

As a learning algorithm, for example, Rumelhart, DE, McClelland, JL and the PDP Re
search Group, PARALLEL DISTRIBUTED PROCESSING, the
Backpropagation described in MIT Press, 1986. can be used.

[0015]

1 is a schematic diagram showing a word recognition device used in a first embodiment of the present invention, FIG. 2 is a schematic diagram showing a word recognition device used in a second embodiment of the present invention, and FIG. 3 is an input. FIG. 4 is a schematic diagram showing a voice division state, FIG. 4 is a schematic diagram showing a neural network, FIG. 5 is a schematic diagram showing a hierarchical neural network, and FIG. 6 is a schematic diagram showing a structure of a unit.

(First Embodiment) (Refer to FIGS. 1 and 3) The word recognition device 10 includes a voice input unit 1 as shown in FIG.
1, a power calculation unit 12, a peak position detection unit 13, an n-channel bandpass filter 14, a block division averaging unit 15, a network selection unit 16, neural networks 17A, 17B, 17C ..., And a determination unit 18. It

(A) Learning 1. Input Creation The registered words are 100 words, and each word voice is input by the voice input unit 1.
Entered in 1. The power is obtained for each frame from the amplitude of each word voice, and the time series of the power for each word voice is obtained. At this time, smoothing processing is performed by averaging a plurality of frames.

Using the peak position detector 13, a frame corresponding to the maximum point (peak) of power is obtained from the power time series.

As shown in FIG. 3, the block division averaging unit 15 is used to distinguish the obtained frame from other frames, and to identify the blocks and peaks of the corresponding frame between the voice start point and the peak. The audio is divided into blocks of corresponding frames, blocks of corresponding frames between peaks and peaks, and blocks of corresponding frames between peaks and end points.

The speech in the divided blocks is passed through an n-channel bandpass filter 14, and the block division averaging unit 15 averages each band to obtain an input of the neural network.

2. Learning The neural network 17A, 17B, 17 having the input unit selected by the network selection unit 16 and having the input unit corresponding to the number of divisions of the input voice
Enter in C ... In addition, the neural network 17A,
17B, 17C ... are prepared for each number of divisions of the registered word, and 1
One network uses an input layer having units of n (channel) x (number of divisions of registered word) and an output layer having units of 100 (pieces).

100 words are numbered and corresponded to 100 units in the output layer, and for the input determined above,
Each network 17A, so that the value of the output unit corresponding to the word is 1 and the value of the other units is 0,
17B, 17C ... are sufficiently learned.

(B) Evaluation 1. Input Creation Using the power calculation unit 12 for the voice collected by the voice input unit 11, the power is calculated for each frame from the amplitude of each word voice, and the time series of the power for each word voice is calculated. Ask.
At this time, smoothing processing is performed by averaging a plurality of frames.

Using the peak position detection unit 13, a frame corresponding to the maximum point (peak) of power is obtained from the power time series.

As shown in FIG. 3, the block division averaging unit 15 is used to distinguish the obtained frame from other frames, and to identify the blocks and peaks of the corresponding frame between the voice start point and the peak. The audio is divided into blocks of corresponding frames, blocks of corresponding frames between peaks and peaks, and blocks of corresponding frames between peaks and end points.

The voices in the divided blocks are passed through an n-channel bandpass filter 14, and the block division averaging unit 15 averages each band to obtain an input of the neural network.

2. Evaluation network selecting section 16 selects learned networks 17A, 17B, 17C ... Corresponding to the number of divisions, and selected neural networks 17A, 17B, 17
Enter the value obtained above in C.

When there is no neural network corresponding to the number of divisions, voice input is prompted again.

The determining unit 18 determines the input word from the values of the units in the output layers of the neural networks 17A, 17B, 17C.

(C) Experiment The following Experiment 1 (conventional method) and Experiment 2 (inventive method) were performed for 100 recognized words (person name) and one specific speaker.

Experiment 1 The input speech is temporally equally divided (8 blocks). The average of the frequency characteristics calculated in each block (using a bandpass filter of 16 channels) was input to the neural network for learning and evaluation.

Experiment 2 The power is calculated for each frame from the input speech, the frames corresponding to a plurality of peaks are calculated from the time series of the power, the obtained frames and the other frames are distinguished, and the start point and the peak of the speech are distinguished. The audio is divided into blocks of a frame corresponding to the intervals, blocks of a frame corresponding to the peaks, blocks of a frame corresponding to the peaks and peaks, and blocks of a frame corresponding to the peaks and end points. Next, the value obtained by calculating the average of the frequency characteristics of the voice in each block was used as the input of the neural network for learning and evaluation.

In Experiment 2 using the method of the present invention, the error rate was about 3/5 as compared with Experiment 1 of the conventional method.

That is, according to the method of the present invention, redundant feature quantities (frequency characteristics) are averaged and replaced by one representative point,
Since the input speech is divided considering the simple phoneme classification,
We confirmed that it is resistant to nonlinear expansion and contraction of speech in the time direction due to differences in vocalization habits among speakers and variations in vocalization within speakers, and can reduce the amount of computation (speed) and improve recognition accuracy at the same time.

(Second Embodiment) (See FIGS. 2 and 3) The word recognition device 20 differs from the word recognition device 10 in that
Only one neural network 17 is used as the neural network.

(A) Learning 1. Input Creation The registered words are 100 words, and each word voice is input by the voice input unit 1.
Entered in 1. The power is obtained for each frame from the amplitude of each word voice, and the time series of the power for each word voice is obtained. At this time, smoothing processing is performed by averaging a plurality of frames.

Using the peak position detector 13, the frame corresponding to the maximum point (peak) of power is obtained from the power time series.

As shown in FIG. 3, the block division averaging unit 15 is used to distinguish the obtained frame from other frames, and to identify the blocks and peaks of the corresponding frame between the start point and the peak of the voice. The audio is divided into blocks of corresponding frames, blocks of corresponding frames between peaks and peaks, and blocks of corresponding frames between peaks and end points.

The voices in the divided blocks are passed through an n-channel bandpass filter 14, and the block division averaging unit 15 averages each band to obtain an input of the neural network.

2. Learning The input obtained above is input to the neural network 17. The neural network 17 uses an input layer having units of n (channel) × (the maximum number of blocks among the number of divisions of registered words) and an output layer having 100 units. If the input units of the neural network are left over, 0
Is substituted.

The 100 words are numbered and corresponded to 100 units in the output layer. For the input obtained above, the value of the output unit corresponding to that word is 1, and the value of the other units is 0. The network 17 is sufficiently learned so that

(B) Evaluation 1. Input Creation Using the power calculation unit 12 for the voice collected by the voice input unit 11, the power is calculated for each frame from the amplitude of each word voice, and the time series of the power for each word voice is obtained. Ask.
At this time, smoothing processing is performed by averaging a plurality of frames.

Using the peak position detector 13, a frame corresponding to the maximum point (peak) of power is obtained from the power time series.

As shown in FIG. 3, the block division averaging unit 15 is used to distinguish the obtained frame from other frames, and to identify the blocks and peaks of the corresponding frame between the start point and the peak of the voice. The audio is divided into blocks of corresponding frames, blocks of corresponding frames between peaks and peaks, and blocks of corresponding frames between peaks and end points.

The voices in the divided blocks are passed through an n-channel bandpass filter 14, and the block division averaging unit 15 averages each band to obtain an input of the neural network.

2. Evaluation The value obtained above is input to the learned network 17. When there are surplus input units in the neural network 17, 0 is substituted into the surplus units. Also,
Prompt for voice input again when there is a shortage.

The judging section 18 judges the input word from the value of each unit in the output layer of the neural network 17.

(C) Experiment The following Experiment 1 (conventional method) and Experiment 2 (inventive method) were performed for 100 recognized words (person name) and one specific speaker.

Experiment 1 The input voice is temporally equally divided (8 blocks). The average of the frequency characteristics calculated in each block (using a bandpass filter of 16 channels) was input to the neural network for learning and evaluation.

Experiment 2 The power is calculated for each frame from the input voice, the frames corresponding to a plurality of peaks are calculated from the power time series, the obtained frame and the other frames are distinguished, and the start point and the peak of the voice are determined. The audio is divided into blocks of a frame corresponding to the intervals, blocks of a frame corresponding to the peaks, blocks of a frame corresponding to the peaks and peaks, and blocks of a frame corresponding to the peaks and end points. Next, the value obtained by calculating the average of the frequency characteristics of the voice in each block was used as the input of the neural network for learning and evaluation.

In Experiment 2 using the method of the present invention, the error rate was about 2/3 of that in Experiment 1 of the conventional method.

That is, according to the method of the present invention, redundant feature quantities (frequency characteristics) are averaged and replaced by one representative point,
Since the input speech is divided considering the simple phoneme classification,
We confirmed that it is resistant to nonlinear expansion and contraction of speech in the time direction due to differences in vocalization habits among speakers and variations in vocalization within speakers, and can reduce the amount of computation (speed) and improve recognition accuracy at the same time.

[0053]

As described above, according to the present invention, it is resistant to non-linear expansion and contraction of voice in the time direction due to differences in vocalization habits among speakers and variations in vocalization within speakers, and reduction in the amount of calculation (speedup). ) And the recognition accuracy can be improved at the same time.

[Brief description of drawings]

FIG. 1 is a schematic diagram showing a word recognition device used in a first embodiment of the present invention.

FIG. 2 is a schematic diagram showing a word recognition device used in a second embodiment of the present invention.

FIG. 3 is a schematic diagram showing a divided state of an input voice.

FIG. 4 is a schematic diagram showing a neural network.

FIG. 5 is a schematic diagram showing a hierarchical neural network.

FIG. 6 is a schematic view showing a structure of a unit.

[Explanation of symbols]

10, 20 word recognition device 11 voice input unit 12 power calculation unit 13 peak position detection unit 14 bandpass filter 15 block division averaging unit 16 network selection unit 17, 17A, 17B, 17C neural network 18 determination unit

Claims (3)

[Claims] 1. In a word recognition method for recognizing a word from an input voice using a neural network, the power is obtained for each frame from the input voice, and the frames corresponding to a plurality of peaks are obtained from the time series of the power. Differentiating between frames and other frames, the block of the frame corresponding to the start point and the peak of the voice, the block of the frame corresponding to the peak, the block of the frame corresponding to the peak to peak, the peak and the end point Divide the audio into each of the blocks of the corresponding frame between and,
A word recognition method characterized in that a value obtained by calculating an average of frequency characteristics of voices in each block is input to a neural network. 2. The word recognition method according to claim 1, wherein a neural network having an input unit corresponding to the number of divisions of input speech is used as the neural network. 3. As the neural network, 1
The word recognition method according to claim 1, wherein two neural networks are used.