JPH06289899A - Speech recognition device - Google Patents

Abstract

PURPOSE:To correspond the speech recognition device which uses a neural network as a recognition part to aging variation in the pronunciation of a speaker with time by letting a neural network part learn speech feature data which causes misrecognition in case of misrecognition. CONSTITUTION:The neural network 3 incorporated in the speech recognition device learns the speech feature data which causes misrecognition automatically at each occasion under the control of a speech recognition and learning control part 7. In order to prevent defective data from being learnt, the learning is performed only when the speech feature data match data which are already learnt to some extent. Therefore, even if the pronunciation of the speaker varies with time, the speech feature data are learnt at each time to adapt the device to the variation in the pronunciation of the speaker, thereby maintaining high recognition precision.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device for recognizing a word voice uttered by a speaker and outputting the result.

[0002]

2. Description of the Related Art In a conventional voice recognition device for recognizing a word voice using a neural network unit, the voice feature data of the word voice is learned in advance in a neural network unit, and the voice feature data is passed to the neural network unit during recognition. Then, the degree of coincidence with the learned word is output as the recognition result.

[0003]

In the conventional voice recognition device, when the pronunciation of the speaker changes with time after learning of the neural network unit, a word recognition error is likely to occur. Then, unless the learning data of the neural network part for recognition is input again and learning is performed again, the recognition error is not improved. In addition, learning bad data including noise and peculiar pronunciation at that time may rather reduce the recognition accuracy.

An object of the present invention is to solve the above conventional problems and to provide a voice recognition device capable of maintaining high recognition accuracy in response to changes in the pronunciation of a speaker.

[0005]

In order to achieve the above object, the present invention provides a voice recognition device using a neural network for recognizing word voices uttered by specific and unspecified speakers,
A voice input unit for inputting a voice signal, a voice signal feature extraction unit for extracting the features from the voice signal, a learning data storage unit for the neural network, a recognition result display, and the correctness received from the user External The interface unit, a neural network unit that outputs the degree of coincidence between the extracted voice feature data and the voice feature data of each learned word by numbers, and learns the given voice feature data, and the neural network unit. It is characterized in that it is provided with a recognition result judging unit for judging a recognition result from the output coincidence degree, and a voice recognition / learning control unit for controlling the flow of the data and for making the neural network unit perform learning.

[0006]

According to the present invention, the neural network unit incorporated in the voice recognition device is made to automatically learn the voice feature data for which recognition has been missed. Further, in order to prevent the learning of defective data, the learning is executed only when the voice feature data matches the learned data to some extent.

Therefore, even if the pronunciation of the speaker changes with time, by learning the voice feature data each time, it is possible to respond to the change in the pronunciation of the speaker and maintain high recognition accuracy.

[0008]

1 is a functional block diagram of a voice recognition apparatus according to an embodiment of the present invention. In FIG. 1, 1 is a voice input section for inputting a voice uttered by a speaker, and 2 is the voice input. A voice signal feature extraction unit for calculating voice feature data from the voice signal input by the unit 1, and a neural network for receiving the voice feature data as an input and outputting the degree of matching with the voice feature data of each learned word. A learning unit 4 receives a recognition result from the neural network unit 3 and calculates a top 3 words with a high degree of coincidence. A learning unit 5 stores voice feature data of the words learned by the neural network unit 3. An external interface unit for receiving a recognition result from the learning data storage unit 5, displaying the recognition result, and allowing the user to input the correctness of the result,
Reference numeral 7 denotes a voice recognition / learning control unit that receives correctness / incorrectness information of the recognition result from the external interface unit 6 and determines whether the neural network unit 3 should perform learning.

FIG. 2 is a circuit block diagram of the voice recognition apparatus of FIG. 1, where 8 is a microphone and 9 is a read-only memory.
(Hereinafter, abbreviated as ROM), 10 is a central processing unit (hereinafter,
Reference numeral 11 is a random access memory (hereinafter abbreviated as RAM), 12 is a monitor, and 13 is a keyboard.

The voice input section 1 shown in FIG. 1 is a microphone 8 and the learning data storage section 5 is a RAM 11 shown in FIG.
The voice signal feature extraction unit 2 and the neural network unit 3
The recognition result determination unit 4 and the voice recognition / learning control unit 7
The external interface unit 6 is realized by the monitor 12 and the keyboard 13 as the U 10 executes the program stored in the ROM 9 while exchanging data with the ROM 9 and the RAM 11.

This operation will be described below in the case where the voice recognition device in the embodiment of the present invention configured as described above is given a voice of the word "tanaka" with a pronunciation different from that at the time of initial learning. Will be described with reference to the flowchart of FIG. It is assumed that the neural network unit 3 has already learned the learning data shown in (Table 1). There are two data for each word, and the size of one data is 27.
It is 0 bytes and consists of 45 numbers.

[0012]

[Table 1]

In step (S1), the voice signal feature extraction unit 2
Performs voice feature extraction for the input voice signal input from the voice input unit 1.

In step (S2), the extracted voice feature data is input to the neural network unit 3, and the degree of coincidence with each learned word is obtained as an output. The recognition result determination unit 4 calculates the top three words in the descending order of the degree of matching. In the above example, as shown in (Table 2), it is assumed that "Tanaka" is in second place and "Tokuda" is in first place.

[0015]

[Table 2]

In step (S3), the external interface unit 6 outputs the recognition result to the monitor 12, and the user looks at it and gives a yes when the recognition result is correct and a correct answer when the recognition result is incorrect. Input on the keyboard 13. In the above example, since the recognition result is incorrect (no), the user inputs “tanaka”.

In step (S4), the voice recognition / learning control unit 7 receives the user's input from the external interface unit 6, and if the input is yes, the process is terminated. When the input is no, if the correct word input by the user is within the third rank for recognition, the process proceeds to step (S5), and if not entered, the process ends. This is to prevent defective data containing noise and data with greatly disturbed pronunciation from being mixed into the learning data. In the above example, since "Tanaka" is in the second place, it is regarded as normal data and the step
Proceed to (S6).

In step (S5), the old voice feature data in the learning data in the learning data storage unit 5 is updated to the voice feature data that was erroneously recognized this time, and is stored in the RAM 11.
In the above example, the older one (No. 3) of the voice feature data of "Tanaka" in the learning data of (Table 1) is erased, and the data of "Tanaka" which is erroneously recognized this time is recorded. In other words, (Table 3) shows the changed learning data, new data is inserted in No. 3, and the original No. 3 becomes No. 4.

[0019]

[Table 3]

In step (S6), the neural network unit 3 is made to learn the updated learning data, and the learned neural net is stored. The initial learning of the neural network part requires several thousand times of learning, but if only part of the learning data is updated, about 100 times of learning is sufficient for the learning to converge, and the time required for learning is within the practical range. And
Experiments have revealed that the learned data can be accurately recognized. In the above example, the neural network unit learns the change in the pronunciation of "Tanaka",
Later, even if the pronunciation of "tanaka" changes, it will be possible to correctly recognize it.

As described above, by adding the learning function of the neural network section and the function of discriminating normal data to the conventional voice recognition device, the device automatically learns only normal data when the user points out an error. , It is possible to maintain high recognition accuracy even if the pronunciation of the speaker changes over time.

[0022]

As described above, the voice recognition apparatus of the present invention can obtain high recognition accuracy by automatically learning the change in the pronunciation of the speaker without requiring any special operation.

[Brief description of drawings]

FIG. 1 is a functional block diagram of a voice synthesizer according to an embodiment of the present invention.

FIG. 2 is a circuit block diagram of FIG.

3 is a flowchart illustrating the operation of FIG.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... Voice input part, 2 ... Voice signal feature extraction part, 3 ... Neural net part, 4 ... Recognition result determination part, 5 ... Learning data storage part, 6 ... External interface part, 7 ... Voice recognition / learning control part, 8 ... Microphone, 9 ... RO
M, 10 ... CPU, 11 ... RAM, 12 ... Monitor,
13 ... keyboard.

Claims (1)

[Claims] 1. A voice recognition apparatus using a neural network for recognizing a word voice uttered by a specific / unspecified speaker, and a voice input unit for inputting a voice signal, and a feature thereof is extracted from the voice signal. The voice signal feature extraction unit, the learning data storage unit of the neural network, the external interface unit that displays the recognition result and receives the correctness from the user, the extracted voice feature data and the voice feature data of each learned word A neural network unit that outputs the degree of coincidence as a number and learns given voice feature data, a recognition result determination unit that determines a recognition result from the degree of coincidence output by the neural network unit, and a flow of the data. A voice recognition device comprising: a voice recognition / learning control unit that controls the neural network unit to perform learning.