JPH08314489A - Voice recognition device - Google Patents

Abstract

PURPOSE: To recognize vocalization of voices simultaneously inputted to plural microphones by a voice recognition device. CONSTITUTION: This voice recognition device is provided with microphones 1-1 to 1-N converting an input voice to a voice signal, input voice analysis parts 2-1 to 2-N analyzing the voice signal from the microphones 1-1 to 1-N and converting to a characteristic vector system, characteristic vector system storage parts 3-1 to 3-N storing the characteristic vector system from the input voice analysis parts 2-1 to 2-N and a characteristic vector system monitor switch part 4 monitoring respective characteristic vector system storage parts 3-1 to 3-N and successively selecting and switching the storage part storing the effective characteristic vector system. Further, the device is provided with a recognition part 6 calculating the similarity between the effective vector system from respective storage parts and a standard pattern beforehand stored in a recognition dictionary part 5 and making the standard pattern that the similarity becomes maximum the recognition result of the input voice.

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, and more particularly to a voice recognition device for deriving a recognition result of an input voice by pattern matching between an input voice pattern and a recognition dictionary.

[0002]

2. Description of the Related Art Conventionally, there has been generalized a system in which a plurality of workers perform a similar work at an industrial site or the like to input information by voice into a computer or the like to process information. In such a system, the contents input by voice are represented by simple words in advance, and the word voice input by each worker is recognized by the voice recognition provided for each worker. A voice recognition method is adopted in which the recognition is performed by the device and the recognition result is sent to the host computer. Then, the host computer appropriately sends an answer to the worker by executing predetermined information processing based on the sent recognition result, and advances a series of work. As a means for the host computer to send an answer to the operator, there are a method of replying to the operator by voice through a response device, a method of displaying the answer on a display device such as a CRT and presenting it to the operator.

Referring to FIG. 2 showing the voice recognition method as the first conventional technique, the voice uttered by each worker is converted into an electric signal in each of the microphones 10-1 to 10-N. Corresponding voice recognition devices 11-1 to 11
-N. The result of voice recognition performed by each of the voice recognition devices 11-1 to 11-N is sent to the host computer 12, and the host computer 12 receives the voice recognition result input from any of the voice recognition devices. CRTs 13-1 to 13-corresponding to each of the speech recognition results by identifying whether there is any, and using the processing result as an answer.
It is sent to each N and displayed on the screen of each N. Therefore, for example, when a voice is input from the microphone 10-1, the answer is displayed on the CRT 13-1 and the worker sees this answer and proceeds with the work.

However, the frequency of occurrence of an event is low depending on the contents of the work handled by the system of the voice recognition system which is the first prior art, and even if a plurality of voice recognition devices are installed, it is possible that voice input occurs simultaneously. Extremely low. Therefore, in some cases, two or more voice recognition devices do not operate at the same time. In such a case, there is a problem that the operating rate of the voice recognition device is extremely low, and it is very uneconomical to install a plurality of expensive voice recognition devices.

Japanese Unexamined Patent Publication No. 1-216398 discloses a speech recognition system as a second conventional technique which improves the problems of the speech recognition system as the first conventional technique.

The second prior art voice recognition system will be described with reference to FIG. 3. The voice recognition system is installed in correspondence with the number of voice input persons, and the uttered voice of each voice input person is converted into an electric signal. Microphones 20-1 to 20- for converting to audio signals
N and a voice output period detecting device 21-1 for detecting the voice output period of each voice signal input from the corresponding microphones 20-1 to 20-N and supplying the detection result to the host computer 24 described later. ˜21-N, and a logical sum (OR) circuit 22 that supplies the speech signals output from the microphones 20-1 to 20-N to a speech recognition device 23 described later and is supplied from the OR circuit 22. A voice signal that recognizes (for example, converts into a corresponding code or the like) and supplies the result to the host computer 24, and a predetermined voice recognition result based on the voice recognition result input from the voice recognition device 23. The host computer 24 that performs processing and displays the processing result on each screen of each of the corresponding CRTs 25-1 to 25-N, which will be described later, and from the host computer 24 Answer to display the above-mentioned CRT25-
1 to 25-N.

Next, the operation of the voice recognition system which is the second conventional technique will be described. First, the operation when there is audio input from only one channel will be described. When an event occurs in the work process, it is assumed that a worker utters a word voice for the event toward the microphone 20-1. The microphone 20-1 electrically converts the collected voice into a voice signal and outputs the voice signal to the voice recognition device 23 via the OR circuit 22. The voice recognition device 23 performs a predetermined recognition process for recognizing an input voice signal, converts it into a code that can be discriminated by a computer, and outputs this code to the host computer 24 as a recognition result of the input voice.
On the other hand, the voice signal output from the microphone 20-1 is input to the voice output period detection device 21-1, and the voice output period, that is, the voice utterance period is detected here.

Audio output period detectors 21-1 to 21-N
The output signal of each respective is input to the host computer 24, and the host computer 24 identifies by which audio output period detection device these output signals are input.
It is known that a voice is input from the voice input channel, that is, the microphone 20-1 in the above case. As a result, when the host computer 24 performs the processing based on the recognition result of the voice recognition device 23, it sends the processing result to the CRT 25-1 corresponding to the voice input channel as an answer,
A response message such as the next work instruction is displayed on the screen of the CRT 25-1.

As described above, when a certain channel is performing audio input, the host computer 24 determines the channel number of the audio input based on the audio output period detection signal output from the corresponding audio output period detection device. can do. Therefore, the voice recognition device 23 sends a message A indicating that the voice recognition device 23 is in use to the CRT of another channel and displays it on the screen. The message A includes, for example, "the voice recognition device is busy. Please wait for a while for voice input". Host computer 24
After that, when the input processing of the channels is completed, for example, a message B "Voice recognition device is ready. Voice input is possible" is sent to the CRTs of all channels.

Next, when a plurality of channels simultaneously input voices, the output of the OR circuit 22 is a superposition of a plurality of voices.
Input to 3. Therefore, even if these voices are recognized by the voice recognition device 23, the recognition result is not reliable. However, in this case, when a plurality of audio output period detection signals that overlap in time are input from the audio output period detection devices of a plurality of channels to the host computer 24, the host computer 24 receives audio input from a plurality of channels at the same time. That is, the recognition result input from the voice recognition device 23 is discarded, and the priority order is determined for each of the plurality of channels that input the voice at the same time. Then, the host computer 24 sends out, for example, a message C "Please generate the voice again" to the CRT of the channel having the highest priority and displays it. At the same time, the host computer 24 sends a message A similar to the above, that is, "the voice recognition device is busy. Please wait for a while for voice input" to the CRT of the channel of another priority. In this way, when the voice input processing of the channel with the highest priority is completed, the host computer 24 sends the message C to the channel with the second highest priority, and at the same time with respect to the channels with other priorities. Performs a process of transmitting the message A. Eventually, the host computer 24 repeats the above-mentioned operation so that the voice input is time-divisionally performed for all the channels to which the voice is input redundantly, and the effective recognition result from the voice recognition device 23 is obtained, and the voice input processing To proceed sequentially. It should be noted that the message regarding the use state of the voice recognition device 23 is displayed in the corner of the screen of the CRT, for example, and the instruction such as work contents is displayed in a large size in the center of the screen.

According to the second prior art voice recognition method, when voices are simultaneously input from a plurality of channels, priority is determined and voices from the respective channels are input in a time division manner. Therefore, only one voice recognition device 23 needs to be provided.

[0012]

In the conventional voice recognition device based on the voice recognition system which is the second conventional technique, when voices are simultaneously uttered by a plurality of microphones,
It was necessary to determine the priority of multiple voice input persons and request them to speak again. For this reason, there is a problem that the voice input person becomes less efficient in work due to waiting for input and re-voice.

[0013]

A speech recognition apparatus according to the present invention comprises input speech analysis means for analyzing a speech word input via a microphone for each utterance and converting it into a feature vector sequence, and the input speech. A plurality of voice analysis storage units having a feature vector storage unit that stores the feature vector sequences from the analysis unit, and an effective feature vector sequence by monitoring each of the feature vector storage units of the plurality of voice analysis storage units. Of the stored feature vector storage means for sequentially selecting and switching the feature vector sequence monitoring switching means, and each of the plurality of voice analysis storage units input via the feature vector sequence monitoring switching means. The degree of similarity between the valid feature vector sequence from the feature vector storage means and the standard pattern of the voice word stored in advance as a recognition dictionary is calculated. Audio word of the standard pattern the similarity is maximized by and a recognition means for the recognition result of the input speech.

Further, in the voice recognition apparatus according to the present invention, when the input voice analysis means finishes analyzing the voice word for each utterance, the addition of notifying that the analysis of the voice word for one utterance is finished It is characterized in that information is added to the feature vector series.

Further, the voice recognition apparatus according to the present invention determines that the feature vector sequence is valid at the time when each of the plurality of input voice analysis means finishes analyzing the voice word for each utterance. Characterize.

Furthermore, in the speech recognition apparatus according to the present invention, each of the plurality of feature vector storage means stores the feature vector sequence of the voice word for each utterance in the memory when the feature vector sequence ends. Is determined to be effective.

[0017]

Next, the present invention will be described with reference to the drawings. Referring to FIG. 1 showing an embodiment of the present invention, a voice recognition device is installed corresponding to the number of a plurality of voice input persons, and electrically converts a voice output from each voice input person into a voice signal. An input voice analysis unit 2-1 that analyzes a voice word for each utterance of a voice signal input from the microphones 1-1 to 1-N and a voice signal input from the microphones 1-1 to 1-N and converts the voice word into a feature vector sequence. 2-N and the input voice analysis unit 2-1.
2-A feature vector sequence storage unit 3 that stores the feature vector sequence input from N in units of voice words for each utterance
1 to 3-N and feature vector sequence storage units 3-1 to 3-N
A feature vector sequence monitoring switching unit 4 for sequentially selecting and switching a storage unit having an effective feature vector sequence among the respective feature vector sequences stored in, and a recognition dictionary for storing standard voice patterns of a plurality of voice words. The voice recognition result of the input voice is obtained by performing pattern matching between the effective voice vector from the feature vector sequence monitor switching unit 4 and the standard voice pattern from the recognition dictionary unit 5 by the unit 5. And a recognition unit 6.

Next, the operation will be described. First, the operation when a voice is input from only one microphone will be described. When the user X utters a word to be recognized by the microphone 1-1, the microphone 1-1 electrically converts the input uttered voice into a voice signal and outputs the voice signal to the input voice analysis section 2-1. Input voice analysis unit 2
-1 analyzes a voice word unit for each utterance of the input voice signal, converts it into a feature vector sequence, and outputs it to the feature vector sequence storage unit 3-1. The feature vector sequence storage unit 3-1 stores the converted feature vector sequence in the memory in units of voice words for each utterance. The feature vector sequence monitoring switching unit 4 is a feature vector sequence storage unit 3-1 to 3
Each of -N is constantly monitored, and, for example, when the feature vector series stored in the feature vector series storage unit 3-1 becomes valid, this output is selected and switched to be connected to the recognition unit 6. When the recognition unit 6 is connected to the output of the feature vector sequence storage unit 1-1 via the feature vector sequence monitoring switching unit 4, the feature vector sequence stored and stored in this memory is stored in the recognition dictionary unit 5 in advance. The similarity of the stored and stored voice word to the standard pattern is calculated, and the voice word of the standard pattern having the maximum pattern similarity is used as the recognition result of the input voice.

Next, the operation when voices are simultaneously input from a plurality of microphones will be described. The voices simultaneously input from the microphones 1-1 to 1-N are converted into respective feature vector sequences by the input voice analysis units 2-1 to 2-N, and feature vector sequence storage units 3-1 to 3-3-
Each of the N memory units stores and stores a voice word unit for each utterance. The feature vector sequence monitor switching unit 4 constantly monitors each of the feature vector sequence storage units 3-1 to 3-N, and a storage unit having an effective feature vector sequence is predetermined among these storage units. The connection is sequentially selected and switched according to the priority order. When the recognition unit 5 is connected to any of the feature vector sequence storage units 3-1 to 3-N via the feature vector sequence monitoring switching unit 4, the recognition unit 5 is similar to the case where there is a voice input from only one microphone described above. When the recognition process is completed, a signal notifying the end of recognition is output to the feature vector sequence monitoring switching unit 4. Then, the feature vector sequence monitoring switching unit 4
Upon receiving a notification from the recognition unit 6 that the first valid feature vector sequence has been recognized, the storage unit having the next valid feature vector sequence is selectively switched according to the next priority so as to be connected to the recognition unit 6. . Hereinafter, the same operation is repeated until the voice input from all microphones is recognized.

As described above, the feature vector sequence monitor switching unit 4 determines that the feature vector sequences in two or more storage units are valid at the same time while monitoring the feature vector sequence storage units 3-1 to 3-N. When this is done, the recognizing unit 6 selects the storage unit with the higher priority, which is predetermined in this system.
Select to connect to and switch.

Here, the criterion for the feature vector sequence monitor switching unit 4 to determine which of the feature vector sequence storage units 3-1 to 3-N is valid is, for example, the input voice analysis unit. This is the time when it is confirmed that the analysis of the voice words for the utterance has been completed and that the feature vector sequence of the voice word for one voice for which the analysis has been completed is stored in the memory in the feature vector sequence storage unit.

As the confirmation method, for example, when the input voice analysis unit finishes analyzing the voice word for each utterance, the additional information notifying that the analysis of the voice word for one utterance is finished is added to the feature vector sequence. Is supplied to the feature vector sequence storage unit and stored in the memory for storage. As a result, the feature vector sequence monitor switching unit 4 determines the feature of the voice word for each utterance from the additional information of each feature vector sequence stored in each of the feature vector sequence storage units 3-1 to 3-N. The validity or invalidity of the vector sequence is determined, and the feature vector sequence storage unit that is determined to be effective earlier in time series is selected and connected so as to be connected to the recognition unit 6. Then, when it is determined that the feature vector series in the plurality of storage units are valid almost simultaneously in time series, the connection between the corresponding storage unit and the recognition unit 6 is switched according to the predetermined priority order as described above. Therefore, as in the second conventional voice recognition method, when a plurality of voices overlap each other on the time axis, only one voice can be recognized, the other voices are discarded, and the discarded voices are discarded. The speaker is not notified of the request for utterance again.

Next, the recognition unit 6 is connected to the feature vector sequence monitor switching unit 4 and the feature vector sequence storage unit 3 is connected.
The similarity between the effective feature vector sequence from each of 1 to 3 to N and the standard pattern stored and stored in the recognition dictionary unit 5 in advance is calculated. As a method of calculating this similarity,
A pattern matching method is used in which a distance value between a feature vector series (input spectrum time series) and a standard pattern (standard spectrum sequence) is obtained by absorbing a change in the time axis direction using a dynamic programming method. Here, the standard spectrum sequence is stored in the recognition dictionary for each recognition word, and as a result of calculating the distance value for all words, the word of the standard spectrum sequence having the minimum distance value is the word of the standard pattern having the highest similarity. It becomes the recognition result as. Since this method of calculating the degree of similarity is a well-known technique in the field of pattern recognition, further description will be omitted.

[0024]

As described above, according to the present invention, even when the utterances of the voices simultaneously input to the plurality of microphones overlap on the time axis, only one voice is recognized and the other voices are discarded, All the input voices can be recognized by one voice recognition device without adopting a means of notifying the utterer of the discarded voices the utterance request again.

[Brief description of drawings]

FIG. 1 is a block diagram showing a voice recognition device according to an embodiment of the present invention.

FIG. 2 is a block diagram showing a first conventional example.

FIG. 3 is a block diagram showing a second conventional example.

[Explanation of symbols]

 1-1 to 1-N Microphone 2-1 to 2-N Input speech analysis unit 3-1 to 3-N Feature vector sequence storage unit 4 Feature vector sequence monitoring switching unit 5 Recognition dictionary unit 6 Recognition unit

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Yasuo Tomooka 5-7-1 Shiba, Minato-ku, Tokyo NEC Corporation

Claims (4)

[Claims] 1. An input voice analysis means for analyzing a voice word for each utterance input via a microphone and converting it into a feature vector series, and a feature vector for storing the feature vector series from the input voice analysis means. A plurality of voice analysis storage units each having a storage unit, and the feature vector storage unit outputs in which effective feature vector series are stored by sequentially monitoring the respective feature vector storage units of the plurality of voice analysis storage units. Feature vector sequence monitor switching means for selecting and switching, and the effective feature vector sequence from each of the feature vector storage means of the plurality of voice analysis storage sections input via the feature vector sequence monitor switching means. And the standard pattern that maximizes the similarity by calculating the similarity with the standard pattern of the speech word stored in advance as a recognition dictionary. Speech recognition apparatus characterized by comprising a recognition means for the speech word and the recognition result of the input speech, the. 2. When the input speech analysis unit finishes analyzing the speech words for each utterance, additional information is added to the feature vector series to inform that the analysis of the speech words for one utterance is finished. Claim 1 characterized by the above.
The voice recognition device described. 3. The feature vector sequence is determined to be valid when each of the plurality of input voice analysis units finishes analyzing the voice word for each utterance. Voice recognition device. 4. Each of the plurality of feature vector storage means determines that the feature vector sequence is valid at the time when the feature vector sequence of the voice word for each utterance has been stored in the memory. The voice recognition device according to claim 1.