JPS59111698A - 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] [Field of Application of the Invention] The present invention relates to a method for improving the speech recognition rate. This invention relates to a speech recognition method in which input speech is correctly recognized by checking the input speech.

[Prior art]

Up until now, speech recognition devices have generally been broadly divided into those that target discrete input voices and those that target continuous input voices, but recently, even if the input voice is discrete, The reality is that speech recognition devices that can recognize even continuous speech are beginning to appear. In such speech recognition devices, several types of standard speech patterns are prepared, one for discrete speech and one for continuous speech, but in the past, the input speech often caused errors. It has become recognized. This means that the speech recognition device can easily determine whether the input speech is discrete or continuous, and the speaker must have taken this into consideration when inputting. This is because it is not.

Therefore, assuming that the discrete input audio is continuous,
Moreover, on the contrary, there was a problem in that continuous input speech was recognized as discrete.

3f) [Object of the Invention] Therefore, an object of the present invention is to provide a speech recognition method that can correctly recognize input speech even if the input speech is discrete or continuous.

[4 Summary of the Invention] For this purpose, the present invention focuses on the fact that speech recognition can be performed more accurately if it is known in advance whether the input speech is discrete or continuous. , confirm in advance whether the input voice after the input voice is discrete or continuous in the initial stage of the service, and use the standard voice pattern selected based on this recognition result to recognize the input voice 7
This is what I decided to do.

[Embodiments of the invention]

The present invention will be explained below with reference to FIGS. 1 and 2.

First, a speech recognition device according to the present invention will be explained. 1st
The figure shows its general configuration.

According to this, the voice is input according to the contents of the instructions emitted from the speaker 2. The control unit 8 issues necessary instructions to the speaker 2 through the voice synthesis unit 7, and the speaker is prompted by the instructions to generate a voice 7 corresponding to the content of the instructions. The uttered voice is first A/D converted by the human voice analysis unit 11 through the microphone 1, and the features of the voice are extracted.The extracted features and P-turns are converted into standard speech by the voice recognition unit 6. By matching patterns, the degree of similarity is obtained as a recognition result. The standard voice patterns to be matched are stored in advance in the standard voice pattern memory 4 by the standard voice/gutter selection section 5 under the control of the control section 8.
However, pattern matching is performed between the standard voice pattern selected in this way. In this case, several types of continuous and discrete standard voice patterns are stored in the duration voice/turn memory 4, but conventionally, pattern matching is performed between all of these standard voice patterns. This meant that I was beginning to fall in love with it. However, in the case of the present invention, since it is confirmed in advance whether the input voice is discrete or continuous, only one of the standard voices and turns is selected. If the input speech is discrete, pattern matching can be performed using a standard speech pattern for discrete speech, and the input speech can be recognized accurately.

By the way, the standard speech pattern includes speech i? corresponding to various words. However, pattern matching will be performed between the speech patterns corresponding to these words. The word with the highest degree of similarity based on these matching results is selected as a candidate for the recognition result, but this situation is the same in pattern matching with other standard speech patterns. Therefore, at the time when the end of pattern matching is reported to the control unit 8, all recognition result candidates corresponding to the selected standard speech pattern have been obtained, but the determination unit 1o selects from among these candidates the ones with the highest degree of similarity. The recognition result is sent to the control unit 8 as the final recognition result. Control unit 8
How this is handled will be described later.

Cf Now, the voice recognition device shown in Figure 1 is used as part of a cash dispenser in banks, etc., and Figure 2 shows the voice recognition processing when used for such purposes. This shows the flow.

As a result, the control unit 8 first instructs the voice input/analysis unit II and the voice recognition unit 6 to prepare for voice input prior to the voice recognition process, and also selects the standard voice pattern for the first input voice to the standard voice pattern selection unit. 5, the standard voice pattern memory 4 is controlled to be read out. In this example, the standard audio pattern for discrete is selected because the first input audio is input discretely in order to specify or distinguish whether the input audio is discrete or continuous. be. Of course, a continuous standard voice pattern may be selected as long as it is easy to recognize. When these preparations are completed, the control unit 8 sends an input reminder message to the speaker 2 through the voice synthesis unit 7 in order to select one of the standard voice patterns for the voice to be emitted. With this message, the speaker inputs voice from microphone 1, but the 1+llll scattered voice input in this case is, for example, J!,
I, if the input audio is continuous, then T(1)
”, and if it is discrete, it is set to “zero (0)”. If the input speech is simple like this, it can be easily recognized using a standard discrete speech pattern. The input voice from the microphone 1 is then subjected to dinotal conversion in the gt loud input analysis section 11, and the characteristics of the input voice are extracted. However, the extracted characteristic pattern of the input voice is then pattern matched with all the discrete (11''l l'T, 14 voice patterns) selected in advance by the voice recognition unit 6. As a result, pattern matching is performed between the speech pattern corresponding to each word included in the standard speech pattern, and a matching result is obtained for the word correspondence.These matching results The single @ with the highest degree of similarity is selected as a candidate for the recognition result, but pattern matching is also performed sequentially with other pre-selected standard speech patterns.Therefore, Control unit 8
At the time when the completion of ni-ya turn matching is reported, all the recognition result candidates corresponding to the selected standard voice/4' turn will be available. The recognition result is sent to the control unit 8 as the final recognition result.

The control unit 8 compares the similarity of the final recognition result from the determination unit 10 with a constant value (rejection constant) to determine whether or not the recognition result is valid. If it is less than a certain value, it is considered invalid and instructs the standard voice i9 turn selection unit 5 to select the same standard voice pattern as before, that is, the standard voice pattern for all discrete use. This prompts you to input your voice. On the other hand, if it is determined that the recognition result is valid, the control unit 8 outputs the voice corresponding to the recognition result and a confirmation request message from the speaker 2 in order to have the speaker confirm whether or not the recognition result is correct. It looks like this. In response to this, pronounceer a inputs the result from microphone 1, but
The confirmation result is not necessarily input by voice.

This is because input may be made from the console section 3 if the console section 3 is provided. As a result, the control section 8
can tell whether the recognition result is correct or not, but if it is a misrecognition, it will be processed in the same way as if the final recognition result was determined to be invalid. On the other hand, if the recognition result is correct, an instruction is given to select a standard speech pattern according to the recognition result. If the recognition result is “row”, select one for continuous, and “zero”
If so, a discrete one is selected, and the recognition process for the input voice is performed using the selected standard voice pattern. This recognition result is sent to the host device 9, thereby completing the speech recognition process for the first input speech, but unless all services are completed, the input speech will be recognized while reminding the next speech input. This is what has become.

〔Effect of the invention〕

0t1 As explained above, in the case of the present invention, it is known in advance from the first input voice whether the input voice will be continuous or discrete, and the voice input thereafter will be either continuous or discrete. Since it can be recognized using a standard speech pattern, there is an effect that the input speech can be identified more reliably.

[Brief explanation of the drawing]

FIG. 1 is a diagram showing a schematic configuration of an example of a speech recognition device according to the present invention, and FIG. 2 is a diagram showing a flow of an example of speech recognition processing in the device. DESCRIPTION OF SYMBOLS 1...Microphone, 4...Standard voice pattern memory, 5...Standard voice/zo turn selection section, 6 speech recognition section, 8...Control section, ]o...Judgment section, 1] ...
Voice input/analysis department,

Claims (1)

[Claims] By analyzing the input voice, characteristic patterns are extracted from the voice, pattern matching is performed between the voice and the standard voice, and from the result of the matching, the input voice YRn is recognized as the voice B method, and the voice is At the beginning of the recognition service, we confirm in advance whether the input voice is continuous or discrete after the input voice, and depending on the confirmation result, we select either a continuous or discrete standard voice pattern. A voice recognition method is characterized in that the voice input after the selection is recognized is performed until the end of the service.