JPH02275497A - Voice recognition device - Google Patents

Abstract

PURPOSE:To improve the recognition rate by performing matching processes in two stages wherein input voices which can be recognized in the respective matching processes are limited respectively. CONSTITUTION:This device is equipped with standard pattern storage means 5 and 6 wherein standard patterns classified in two groups are stored by the groups and 1st stage and 2nd stage matching means which perform similarity calculation. Then this device is provided with a control means 9 which performs return control from the matching process of the 2nd matching means to the matching process of the 1st stage matching means while input patterns in an input pattern storage means at this time are held when the maximum similarity value obtained by the 2nd stage matching process is smaller than a specific threshold value. Therefore, if misrecognition is caused in the 1st step, the voice recognition in the 1st step can be performed even in the 2nd step without moving back to the 1st step from the 2nd step. Consequently, the high-performance recognition rate can be obtained.

Description

DETAILED DESCRIPTION OF THE INVENTION (A) Field of Industrial Application The present invention relates to a voice recognition device capable of recognizing voice and controlling target electrical equipment.

(b) Conventional technology In recent years, with the improvement in the speech recognition rate of speech recognition devices,
Electronic devices that can be controlled by voice, such as telephones that can auto-dial, are being put into practical use (Japanese Patent Laid-Open No. 62-81152).
issue).

For example, in the case of a voice recognition auto-dial telephone, the voice recognition device uses the dial destination name (
The most practical method is one that adopts a two-step recognition processing method in which the user's name (individual name, company name, etc.) is voice-recognized, and the command voice (dial, cancel stop, etc.) is voice-recognized in the second step.

In other words, the two-step recognition process recognizes the name of the dialed destination uttered by the speaker, and instead of immediately dialing the first number of the dialed destination, this recognition result is displayed or confirmed to the speaker using synthesized voice output. Since a dialing command is given by voice when there is no erroneous recognition, it is possible to prevent dialing errors caused by erroneous voice recognition.

A speech recognition device using a pattern matching method that performs such a two-step recognition process uses these words in the first step to recognize the speech of a word that is required only for that step, such as the speech of multiple dialed destination names. A standard group of voice patterns (first group) for the dial destination name is used, and in the second step, a standard group of voice command patterns (second group) are used to recognize multiple command voices. .

These standard patterns of the first group and the second group can be used as one group for the speech recognition processing of both the first and second steps described above, without classifying them, but in this case, The amount of processing required for pattern matching (calculating the error between the input speech pattern and the standard pattern) for speech recognition in each step becomes large, and the probability of misrecognition occurring increases.To avoid this, As mentioned above, the standard speech patterns can be divided into groups corresponding to the recognition processing for each step.

(c) Problems to be Solved by the Invention As mentioned above, in conventional speech recognition devices that perform speech recognition processing in multiple steps, for example, in an auto-dial telephone, the speaker utters a voice in the first step. If the recognition result of the dialed destination name is an incorrect recognition, in the next second step, the command voice 1 cancel uttered by the speaker is recognized, and on the condition that this recognition is performed correctly, the process returns to the first step. It is then necessary to re-speak the name of the dialed destination so that it can be recognized. Such a procedure for re-inputting voice is extremely complicated and has the drawback of increasing the burden on the speaker who inputs voice.

The present invention has been made in order to eliminate the drawback of refusing.If there is an erroneous recognition in the first step, the voice recognition of the first step is repeated at the second step without returning from the second step to the first step. This realizes a speech recognition device that enables processing.

(d) Means for Solving the Problems The speech recognition device of the present invention includes a standard pattern storage means storing standard patterns classified into at least two groups for each group, and an input pattern of the latest input speech. an input pattern storage means, a first stage matching means for calculating similarity by comparing each standard pattern of the first group of the standard pattern storage means and the input voice pattern of the input pattern storage means by pattern matching; a second stage matching means for calculating similarity by comparing each standard pattern of the second group of the standard pattern storage means and the input voice pattern of the input pattern storage means by pattern matching;
Notifying means for notifying the voice recognized by the first stage matching means; if the maximum similarity value obtained in the second stage matching process is smaller than a predetermined threshold, the input pattern at this time is stored in the input pattern storage means; In this state, the apparatus includes a control means for controlling the matching process of the second matching means to be returned to the matching process of the first stage matching means.

(E) Function The speech recognition device of the present invention is a device capable of performing at least two stages of matching processing in which the input weights that can be recognized in each stage of matching processing are each limited, and the speech recognized in the first stage matching is incorrect. When it is determined that the erroneously recognized speech is recognized, by re-inputting the misrecognized speech, the maximum similarity value obtained in the next two-stage matching process becomes smaller than the predetermined threshold. With the pattern saved, it is possible to automatically return from the second stage matching process to the preceding first stage mapping process.

(f) Example FIG. 1 shows the configuration of a speech recognition device according to the present invention.

The device of the present invention shown in the figure includes an input section (1) for inputting audio;
Preprocessing unit (2) that extracts feature parameters from input audio
and standard patterns of two types of groups created in advance,
A standard pattern (5) for matching with the first input audio, a standard pattern (6) for matching with the second input audio, and a preprocessing unit (
The distance (
It is equal to error and has a reciprocal relationship with similarity), and calculate the minimum distance pattern! The basic configuration is an identification section (4) that outputs the identification result.

Furthermore, the configuration of the embodiment shown in the figure will be described in detail below.

First, when the first voice is input to the input unit (1), the preprocessing unit (2) extracts feature parameters from the input voice, and the distance calculation between the input voice and the standard pattern (5) is performed by the identification unit (4). The recognition results are stored in the result storage section (7). Therefore, the control unit (9) waits for a second voice input for a voice command, and when the second voice is input to the input unit (1), the preprocessing unit (2) extracts the feature parameters and then waits for a second voice input. , the characteristic pattern is stored in the input pattern storage section (3). Distance calculation with the standard pattern (6) is then performed in the identification section (4), and the matching distance at that time is compared with a predetermined threshold value in the distance comparison section (8). In other words, the smaller the distance, the greater the similarity. Therefore, if this distance is larger than a predetermined threshold (if the degree of similarity is smaller than a predetermined value), a recognition result invalidation signal is sent to the control unit (9), and the control unit that receives the signal outputs a recognition result invalidation signal. (9) outputs a deletion signal to the result storage unit (7) to delete the recognition result for the first input voice.Furthermore, outputs an input pattern output signal to the input pattern storage unit (3), Identification part (
In step 4), distance calculation is performed between the standard pattern (5) and the result is stored in the result storage section (7).

On the other hand, if it is smaller than the threshold value (if the similarity is larger than a predetermined value), the recognition result of the second voice is sent to the control unit (9).
and outputs a corresponding control signal.

Next, an example in which the speech recognition device of the present invention is employed in an auto-dial T-phone will be shown below.

First, as the 1st group standard pattern of the 1st standard pattern memory (5), the following thorny dial destination names, 2nd N4
The two-word voice commands shown in the table below are prepared as the second group standard patterns in the semi-pattern memory (6).

Table: Standard patterned voice Note that one of the voice commands in the table above is [Send ko (sends the dial that corresponds to the name of the dial destination)], [Send the dial that corresponds to the name of the dial destination], [Send the dial that corresponds to the dial destination name], [Send the dial that corresponds to the name of the dial destination], [Send the dial that corresponds to the name of the dial destination], [Send the dial that corresponds to the name of the dial destination], output).

First, in the first step, the speaker hears the first voice (the other party's voice,
(i.e. the name of the dialed destination), say "Sanyou" and dial the number 1 above.
If the result of recognition processing (distance calculation) using the first group standard pattern in the standard pattern memory (5) is incorrect as "1 matsushita", in the next step the second voice (voice command) 1
All you have to do is say ``candidate'' and output the next candidate through recognition processing using the first group standard pattern of the second standard pattern memory (6), but it is difficult to find the desired one among the lower candidates. It may not come out. In such a case, for example, it is troublesome to return to the first step with an instruction such as 1 cancel, enter the first voice input standby state, and then speak the name of the dialed destination again, as in the conventional device.

In the present invention, when the user suddenly utters ``1 sanyo'' while waiting for the voice command input in the second step, the input voice pattern is stored in the input pattern storage section (3) and Matching is performed between the pattern and the [sending pattern] and [candidate pattern] of the second group standard pattern, respectively, but since the matching distance is larger than a predetermined value, the target standard pattern is 2 From the 2nd group standard pattern of standard pattern memory (6)'
Switch to the second Il quasi-pattern in the $ljM quasi-pattern memory (5), perform matching again, and output the result. In this case, since there are only two target words for the voice command, the threshold value may be set severely.

(G) Effects of the Invention According to the speech recognition device of the present invention, by grouping standard patterns according to recognition target words, it is possible to obtain a high-performance recognition rate even if the threshold at the time of recognition is made strict. can. Furthermore, when performing two-step control of recognition for word selection and recognition for voice commands, there is no need to perform voice commands to select in advance which of the different standard patterns to match. The burden on the user can be reduced.

[Brief explanation of drawings]

FIG. 1 is a block diagram of the speech recognition device of the present invention. (1)...Input section, (2)...Preprocessing section, (3...
- Input pattern storage section, (4)...Identification section, (5...
First standard pattern memory, (6... Second standard pattern memory, (7... Result storage section, (8)... Distance comparison section, (9
)...Control unit.

Claims (1)

[Claims] (1) In a speech recognition device that compares an input speech pattern with a large number of pre-prepared standard speech patterns and recognizes the most similar standard pattern of speech as the input speech, the speech recognition device classifies the input speech into at least two groups. Standard pattern storage means for storing standard patterns in groups; input pattern storage means for storing the input pattern of the latest input voice; and each standard pattern of the standard pattern storage means of the first group and the input pattern storage means. a first stage matching means for calculating the similarity by comparing the input speech patterns of the input speech pattern of the second group with the input speech pattern of the input pattern storage means; a second stage matching means for comparing and calculating similarity through matching; a notification means for notifying the voice recognized by the first stage matching means; a maximum similarity value obtained in the second stage matching processing is a predetermined threshold; control means for controlling the matching processing of the second-stage matching means to be returned to the matching processing of the first-stage matching means with the input pattern stored in the input pattern storage means at this time being smaller than that of the first-stage matching means; A voice recognition device.