JPS5917595A - Voice recognition system - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention aims to improve the efficiency of correction processing in the case of erroneous recognition in a speech recognition device that incorporates (stores and stores) a plurality of sets of standard speech buttons corresponding to each word to be recognized. This relates to a voice recognition method to improve the quality of speech.

The conventional speech recognition method for this type of speech recognition device is
For example, until a series of services is completed, voice recognition processing is performed using multiple sets of built-in standard voice buttons, so a specific voice recognition process by a specific speaker or a general speaker can be If a word is likely to be misrecognized with respect to a certain standard speech pattern, there is a risk that misrecognitions will be concentrated or occur frequently for the word in a series of speech recognition processes.

Furthermore, when an erroneous recognition occurs, the same content is re-uttered, and the speech recognition process is performed with exactly the same content as that in the erroneous recognition.

However, the fact that it was misrecognized means that the target of the misrecognition is the vocalized sound bat or the standard sound bat that should be recognized and judged as being truly close to the above vocalized sound bat. This means that it was closer to the standard speech pattern that became .

Therefore, even if the speech recognition process for the same content is repeated as described above, there is a high probability that the same erroneous recognition will occur repeatedly.
Since a considerable number of vocalizations have to be repeated before a correct recognition result is obtained, there is a problem in that the time required for recognition becomes long and the burden on the speaker increases.

SUMMARY OF THE INVENTION An object of the present invention is to provide a speech recognition method that overcomes the above-mentioned drawbacks of the prior art and, in particular, makes it possible to improve the efficiency of correction processing in the case of erroneous recognition.

The structure of the speech recognition method according to the present invention is to store a plurality of sets of standard speech pattern data corresponding to each word to be recognized, extract the features of the input speech, and combine the feature data with all the above-mentioned standard speech pattern data. In a speech recognition device that has a function of calculating the similarity between the input speech and determining and outputting the standard speech pattern with the highest degree of similarity as the recognition result, it is possible to distinguish between words that are likely to be misrecognized in input speech and the original word. By preparing a correspondence table, when a human voice is misrecognized, control and processing is performed such that a message requesting confirmation of the original word is output based on the correspondence table.

To give a supplementary explanation to this, there are words that have exactly the same acoustic properties (homophones), words that are very similar or similar, and words that are due to distortions, acoustic fabrics, changes, etc. Things that are the same or very similar/similar, etc.
Since there are some words that are easily recognized incorrectly, the correspondence relationships between these words are determined experimentally and statistically, and when a misrecognition occurs, the corresponding word is output for confirmation without having to perform a new recognition process. This is an attempt to improve the efficiency of correction processing. By the way, as a correspondence relationship between words that are easy to make mistakes, for example, 1
(ichi), 8 (stingy) and 1 (ichi).

A typical example is the mutual relationship between 2 and 2.

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 is a system configuration diagram of an embodiment of the speech recognition method according to the present invention, and FIG. 2 is a processing flowchart thereof.

Here, 1 is a standard voice button memory that stores (memorizes) multiple sets of standard voice button data for each word to be recognized, 2 is a standard voice button selection unit that controls the selection, and 3 is a voice button memory. 4 is an analysis unit that extracts features of the human voice; 5 is a speech recognition unit that performs similarity calculation processing (bang matching process) between the feature data and standard voice bang data; 6 is a determination unit that determines a set of standard voice bangs that have a high degree of similarity to the input voice based on the processing results; 7 is a speech synthesis unit that displays the recognition results; 8 is the same speaker; 9 is a unit for displaying the recognition results. Console unit for confirmation and repeated voice input instructions;
Reference numeral 10 denotes a control unit that performs control of the above-mentioned units and other necessary processing, and 11 a host device that performs desired processing based on the recognition result.

1. Prior to recognizing the input voice from the microphone 3, the control unit 10 makes preparations for the voice input in the analysis unit 4. It instructs the speech recognition section 5 and also instructs the standard speech button selection section 2 to select the standard speech button data of the word to be recognized at that time from the standard speech button memory 1 (process 21 in FIG. 2). ).

When these preparations are completed, the voice synthesizing section 7 is instructed to output a human power reminder message to urge the speaker to use voice power, and the message is emitted from the speaker 8 (process 22).

As a result, when the speaker makes a voice input from the microphone 3 (process 23), the analysis unit 4 analyzes the input voice and extracts features (process 24).

As described above, the speech recognition section 5 calculates the degree of similarity between the data of all sets of standard speech slams selected and instructed by the standard speech slam selection section 2 and the characteristic data of the human speech under the control of the control section 1o. Processing (All recognition results and degrees of similarity are conveyed to the determination unit 6 along with the slam pine (processing 25).

The determination unit 6 determines the set of standard speech patterns with the highest degree of similarity, and reports the data to the control unit 1o (process 26).
.

The control unit 10 determines whether the similarity of the recognition result is lower than a predetermined constant (reject constant) and should be considered suspicious to be output as a recognition result (reject). ), in the case of Rizieku F., the process 32) instructs the standard voice button selection section 2 to select the same standard voice button.
Furthermore, it instructs the voice synthesis section to again output a manual reminder message with the same content, and causes the same message to be emitted from the speaker 8 (processing 33). As a result, re-recognition processing (repeated recognition processing) similar to processing 23 and later described above is performed.
will be held.

On the other hand, if the recognition result is not rejected, a confirmation request message is emitted from the speaker 8 via the speech synthesis section 7 as a display to prompt the speaker to confirm whether or not the recognition result candidate is correct (process 28). ).

The speaker listens to this and knows whether the input voice has been correctly recognized (correct recognition) or incorrectly recognized (misrecognition), and inputs this information from the console section 9 to the control section 10. (Process 29).

This manual confirmation of the correctness of the recognition result does not necessarily have to be done by operating the console section 9;
It may be possible to use a human voice for confirmation, but the content is as follows:
It is desirable to have something simple and difficult to misrecognize so that voice recognition can be performed reliably.

The control unit 10 determines whether or not the recognition is correct based on the confirmation information (30), and if the recognition is correct, the control unit 10 transmits the voice slam information to the host device 11 according to the confirmation information of the recognition result as necessary. At the same time, it is determined whether the series of service operations has been completed or not (31).
If the recognition process has not been completed, the process returns to step 21 described above to perform the recognition process for the next input voice, and if the recognition process has completed, the entire recognition result is sent to the host device 11 and the recognition process for one input voice is performed. and prepare for the next manpower.

On the other hand, when confirmation information indicating that the recognition was incorrect is received, a correspondence table (for example, prepared in advance within the control unit 10 itself) of words that are likely to be incorrectly recognized is used to
The original word information is extracted and the speech is immediately emitted from the speaker 8 via the speech synthesis unit 7 as a recognition result (
Processing 34).

The speaker listens to it and manually confirms whether the result is correct or not to the control unit 10 (processing 35).
.

Based on this confirmation, the control unit 10 determines whether the recognition was correct or not (36), and if the recognition is incorrect, performs the process 32.33 in the same manner as in the case of rejection, and if the recognition is correct, makes the above-mentioned judgment. Processing similar to 31 and subsequent steps is performed.

In this way, even if an erroneous recognition occurs, there is no need to perform the recognition process again from the beginning, and in many cases, confirmation and correction can be performed simply by outputting the erroneous word information, so that the correction process can be corrected. Efficiency can be improved.

As described in detail above, according to the present invention, it is possible to improve the efficiency of correction processing when an erroneous recognition occurs, and (9) further improve the overall recognition rate. This will have a significant effect on improving reliability, serviceability, and efficiency in voice recognition systems.

[Brief explanation of drawings]

FIG. 1 is a system configuration diagram of an embodiment of the speech recognition method according to the present invention, and FIG. 2 is a processing flowchart thereof. ■...Standard voice button memory, 2...Standard voice button selection section, 3...Microphone, 4...Analysis section,
5... Speech recognition section, 6... Judgment section, 7... Speech synthesis section, 8... Speaker, 9... Console section, 1o
. . . Control N Department, 11 . . . Host device. Agent Patent attorney Kosaku Fukuda 1 ((1 other person) (10) Kaya Mehaya 2

Claims (1)

[Claims] 1. Store multiple sets of standard voice button data corresponding to each word to be recognized, extract features of human voice,
In a speech recognition device that has a function of calculating the similarity between the feature data and all the above-mentioned standard speech slam data, and determining and outputting the standard speech stamp with the highest degree of similarity as a recognition result, the input speech is incorrectly recognized. By preparing a table of the correspondence between words that are likely to become the same as the original word, if the input speech is misrecognized, a message requesting confirmation of the original word will be output based on the table of correspondence. A voice recognition method that is characterized by control and processing.