JPH0421880B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a speech recognition device, and particularly to an improvement in a device for recognizing continuous word speech uttered by an unspecified speaker.

Conventionally, methods for recognizing speech uttered by unspecified speakers include a learning method that examines the characteristics of the input speech and transforms a standard pattern to match the characteristics, and conversely a learning method that transforms the input speech to match the standard pattern. A normalization method that predicts the range of speech deformation caused by different speakers, and a multi-standard method that places a large number of standard patterns within that range of variation, and a discriminant function method that combines with an appropriate preprocessing method. is proposed. Among these, the ones that currently have practical level recognition ability are based on the multi-standard method and the discriminant function method. Furthermore, considering extending the ability to continuous word recognition, the multi-standard method is almost the only realistic method.

However, in continuous word recognition, considering the possible combinations of word sequences, two-stage DP
Even if methods such as the method or continuous DP method are used, the amount of processing required for recognition increases significantly. Therefore, in a system that performs speaker-independent continuous word recognition while maintaining the multi-standard system, the scale of the pattern matching section etc. becomes extremely large, making it impractical from an economic point of view.

The present invention aims to improve such problems.

Those who are inputting data to a speech recognition device should note that, if they pay attention to a certain usage situation, changes such as changing the gender between men and women or changing from an adult to a child cannot occur during the usage.

That is, in the present invention, the state of the recognition device is divided into a recognition state of discrete utterances (state 1), which requires a relatively small amount of recognition processing, and a recognition state of continuous utterances, which requires a large amount of processing (state 2). The idea is to reduce the processing amount in state 2 by recognizing the input speech in step 1 and limiting the personality of the speaker, and then performing recognition in state 2 using a standard pattern of a common set of characteristics. be.

The present invention will be described below based on Examples.

FIG. 1 shows an example of the configuration of a telephone information service system to which the present invention is applied. It consists of a system control section 1, a voice recognition section 2, a voice response section 3, a hybrid coil 4, and a subscriber telephone 5 according to the present invention, and only the parts necessary for explaining the present invention are extracted from the telephone information service system and written. .

FIG. 2 is a configuration example of a speech recognition device (speech recognition unit 2) for explaining the present invention. In FIG. 2, a control section 21 not only sends and receives commands and results 27 from the system control section 1 of FIG. 1, but also controls the voice recognition section 2. The similarity calculation unit 23 calculates the similarity between the input voice analyzed by the analysis unit 22 and the standard pattern data in the standard pattern memory 24, and the continuous pattern matching unit 25 calculates the optimum matching value between each standard pattern. It is calculated by The result is determined by the determination section 26, and the determination result is sent to the control section 21. The configuration of a recognition device that performs continuous pattern matching processing is already known (Japanese Patent Laid-Open No. 55-2205).
(Refer to the publication) The explanation thereof will be omitted. In the example of this device, the input pattern is always compared with the specified standard pattern, so if it is known in advance that the input is a discrete utterance, the output of the matching section will be the same as that of the input discrete utterance word. In the case of continuous word input, the continuous pattern matching section 25 operates in the same way. The following explanation can be applied in exactly the same manner in a system in which the operation of the matching section 25 is switched between discrete vocalization and continuous vocalization (see, for example, Japanese Patent Application No. 158296/1982).

If the type of word currently registered is "Yes",
“No” and a number from 0 to 9. In addition, the standard pattern of each word and number takes into account differences depending on the speaker, /man / woman / child / 5 types each, i.e. 3 × 5 =
It is assumed that 15 items are registered each. Taking the example of a balance inquiry at a bank, returning to Figure 1, when a call from a user enters the system, the voice response unit 3 first asks the user, ``Do you want to inquire about your balance?'' and the voice recognition unit 2 waits for input in a mode (state 1) in which two types of words, "yes" and "no", are recognized as discrete inputs based on a command from the system control unit 1. When the user answers "yes" or "no," the speech recognition unit 2 only needs to match the two words "yes" and "no" with a total of 30 standard patterns, 15 each. As a result, the standard pattern with the best matching is the group of men (or the group of women, or the group of children).
If so, then in state 2 (continuous word recognition state), the control unit 21 issues a control command to use only the standard pattern of numbers belonging to men (or women, or children). In the next step, the voice response device (voice response section 3) outputs a voice message to the user saying "please enter your password," and the voice recognition section 2 changes to state 2, allowing continuous digit recognition. For example, when the user inputs a password such as "1234" by voice, the voice recognition unit 2 only needs to match it with 10 x 5 = 50 standard numbers patterns belonging to the male (or female, or child) group. It turns out. Therefore, the matching ability of the speech recognition unit 2 is sufficient to match at most 50 standard patterns. On the other hand, in the case of recognition without determining the set in state 1, comparison with 10×15=150 standard patterns is required.

As described above, according to the present invention, it is possible to economically realize a system for recognizing continuous voices uttered by an unspecified speaker, and the effects thereof are significant.

[Brief explanation of drawings]

FIG. 1 shows an example of the configuration of a telephone information service system to which the present invention is applied, and FIG. 2 shows a block configuration of a speech recognition device according to the present invention. 25: Continuous pattern matching section.

Claims (1)

[Scope of Claims] 1. Means for storing a plurality of sets of standard patterns having the same meaning in correspondence with speakers with different characteristics; and means for comparing input speech with the standard patterns. and controlling the matching means to match the input voice of the discrete utterance that is first obtained in a series of responses to an unspecified speaker with the plurality of sets of standard patterns stored in the storage means, The characteristic of the speaker is specified in accordance with the result of the matching means, and one set of standard patterns among the plurality of sets of standard patterns stored in the storage means is limited, and A speech recognition device comprising: means for controlling the collating means to collate continuous utterance input speech with the limited set of standard patterns.