JPS63298298A - Voice section detecting system for voice recognition equipment - 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 Industrial Application) The present invention relates to a speech interval detection method in a speech recognition device.

(Prior Art) In recent years, with the development of voice recognition technology targeting unspecified speakers, voice recognition devices that perform catalog shopping, bank balance inquiries, etc. by recognizing telephone voices have become popular.

The method for detecting a voice section in telephone voice input is different from the method for detecting a voice section in microphone input of a voice input word processor, etc., and this is due to the form of noise other than voiced voice.

Voice input In word processors and the like, voice input is performed using a microphone, which is usually wrapped in a windproof sponge or net. This is to prevent wind noise and noise caused by exhalation from the mouth. However, in the case of a telephone, such measures are not taken, and since noise is generated by holding the transmitter directly in the hand, voice section detection requires more advanced technology than microphone input. .

Furthermore, in the recognition of telephone speech, since the speaker cannot directly check the recognition result by looking at it with his or her mouth, it is essential to check the recognition result using synthesized speech or the like. However, it is unreasonable to check the recognition results based on the voice input during confirmation.
It goes without saying that highly reliable recognition is required for voice input for confirmation.

Therefore, the control words used for confirmation etc. are generally selected from words such as "yes" and "no" that are less likely to cause misrecognition.

(Problems to be Solved by the Invention) However, in the speech section detection method in the conventional telephone speech recognition device, even in recognition that requires a high recognition rate, such as confirmation of recognition results, Since the same speech interval detection is performed, the occurrence of false detections is the same as in the case of normal recognition target words, and there is a problem that highly reliable recognition cannot be performed.

The present invention eliminates the above-mentioned problem of not being able to detect speech segments according to recognition target words, and provides a speech interval detection method that improves the recognition rate of recognition target words that particularly require highly reliable recognition. The purpose is to provide.

(Means for Solving the Problems) In order to solve the above-mentioned problems, the present invention receives audio signals of a plurality of recognition target words prepared for acceptance in advance in accordance with a predetermined order designation, and calculates the audio sections of the audio signals. In the speech section detection method of a speech recognition device, which performs speech recognition by comparing the signal of the speech section with a standard pattern signal, the speech signal of each recognition target word includes the characteristics due to the difference in the number of speech blocks, and the speech signal of each recognition target word is detected. A speech section detection section for each group receives the speech signals of each recognition target word divided into groups based on common features and detects the speech section, and a speech section detection section for each group detects the speech section. and a speech section detection and selection section that selects and distributes the speech sections, and the speech section detection and selection section is selected and designated based on the scheduled order designation.

(Function) According to the present invention, the speech recognition device receives each speech signal according to the scheduled order designation, and the speech section detection and selection section receives each speech signal according to the selection designation based on the order designation. Each voice section detecting section is selected, and each voice signal is sent to each selected voice section detecting section, and the corresponding voice section is detected.

(Embodiment) FIG. 1 is a block diagram of a circuit of a telephone voice recognition device showing an embodiment of the present invention.

In the figure, 1 is a telephone voice recognition device.

Reference numeral 2 denotes an audio input unit which receives an audio signal from a speaker from the telephone line 3.

Although not shown in the figure, when inputting voice, the speaker is sequentially given guidance using synthesized voices from the host device of this telephone voice recognition device 1, and the speaker follows the guidance to input numbers such as the numbers of the code number. A recognition target word that has been prepared in advance is input by voice, and the number etc. is conveyed to the speaker using a synthesized voice for confirmation, and the speaker confirms it by saying "yes" or "no." ” etc. is input.

4.5.6 are first, second, and third voice section detection units, respectively, which detect one or more voice sections according to the selection described later for each voice signal received by the voice input unit 2. Then, the audio section of each audio signal is detected.

FIG. 2 shows the above-mentioned section 1. Second. Third each voice section detection unit 4.
5.6 is an explanatory diagram of voice section detection in section 5.6. In detecting the voice section of the voice signal A, the threshold values LS and LE for the voice block start end and the voice block end are set in advance in each voice zone detecting section 4. 5. If the threshold LS, which is set every 6 minutes, is exceeded continuously for a start end determination time of 78 or more, the first time the threshold LS is exceeded is set as the start of the audio block, and the threshold LE is continued for an end end determination time of TE. If the threshold value LE
The voice block is determined to be the end of the voice block at the time when the voice block first passes, and the voice section is determined from the length of the voice block.

The audio signal may consist of one audio block or multiple audio blocks, and generally includes noise.

Figure 3 is an audio waveform diagram of each vocalization, etc. (1)
is a waveform diagram when the generated sound is "yes", (2) is a waveform diagram when the generated sound is "no", (3) is a waveform diagram when "no" is accompanied by noise, and (4) is a waveform diagram when the sound is "ichi". It is. In speech recognition, it is detected that the waveforms (1) and (2) in the same figure each consist of a speech section made up of one speech block, and the waveform (3) is detected as noise speech. After removing the block, it is necessary to similarly detect that a speech section is composed of one speech block, and for the waveform (4), it is necessary to detect that a speech section is composed of two speech blocks. There is.

The first speech section detection section 4 detects the speech section of the speech signal of a group consisting of only one speech block among each recognition target word, for example, and the second speech section detection section 6 detects the speech section of the speech signal of a group consisting of only one speech block among each recognition target word. However, the third speech section detecting unit 7 detects the speech section of the speech signal for a group in which the number of speech blocks is one or two, and for a group in which there is no limit to the number of speech blocks.

Reference numeral 7 denotes an analysis section which receives the audio signal from the audio input section 2, analyzes the audio signal, and converts the features into parameters. Reference numeral 8 denotes a standard pattern section which stores each pattern of the audio signal prepared for acceptance in advance. Reference numeral 9 denotes a recognition matching section which receives the speech section detection signals from each speech section detection section 4, 5.6 and the parameterized signal from the analysis section 7, and calculates the parameterized signal for the detected speech section. is compared with the pattern in the standard pattern section 8, and if they match, the pattern signal is sent out. Reference numeral 10 denotes a control unit that receives the matching result of the recognition matching unit 9, makes a speech recognition determination, and sends the recognition result to the host device.

Reference numeral 11 denotes a selection designation unit, and as mentioned above, the speaker is sequentially provided with guidance for voice input from the host device.
Since each guidance response can be classified into recognition target word groups with unique properties or recognition target word groups with mixed characteristics, the audio signal input to the audio input unit 2 can be adapted to correspond to each guidance. A selection designation signal is output to be distributed to each voice section detection section 4, 5.6.

Reference numeral 12 denotes a voice section detection and selection section, which selects one or more of the voice section detection sections 4, 5, and 6 based on the selection designation signal of the selection specification section 11, and sends an enable signal to the voice section detection section. Send.

In the case of checking recognition results, etc., the recognition target word is usually
"Yes" or "No", and as shown in FIG. 3, if noise etc. are removed, the voice block consists of one voice block, so the first voice zone detection unit 4
You only need to select.

Since the first voice section detection unit 4 limits the number of voice blocks to only one, if two or more voice blocks are detected due to ambient noise, exhalation, etc., the voice signal is invalidated or the voice waveform is Just select the one closest to you.

Next, the operation of the circuit shown in FIG. 1 will be explained.

The selection designation unit 11 instructs the voice segment detection and selection unit 12, each time a host device (not shown) provides voice input instruction guidance to the speaker, prior to the recognition operation of the corresponding voice input. A selection designation signal is sent out to designate each voice section detection section 4.5.6. The voice section detection and selection section 12 sends out a voice section detection enable signal to one or more of the voice section detection sections 4, 5, and 6. A voice signal input from the telephone line 3 passes through the voice input section 2 and enters each voice section detection section 4.5.6. When the first speech section detection section 4 is selected, only the speech section detection result of the first speech section detection section 4 is input to the recognition matching section 9. On the other hand, the voice input from the voice input section 2 is analyzed by the analysis section 7, and the features are converted into parameters and input to the recognition matching section 9. The recognition matching section 9 performs matching processing with the pattern of the standard pattern section 8 based on the voice section detection result obtained previously, and sends the matching result to the control section 10. The control unit 10 performs recognition determination based on the matching results, and sends the recognition results to the host device.

(Effects of the Invention) As explained above, according to the present invention, the characteristics of the recognition target word are divided into groups and the speech section detection is selected for each group, so that particularly high reliability can be achieved. By selecting appropriate speech segment detection for recognition target words for required confirmation, it is expected that the accuracy of speech recognition will improve.

[Brief explanation of drawings]

FIG. 1 is a block diagram of a circuit of a telephone speech recognition device showing an embodiment of the present invention, FIG. 2 is an explanatory diagram of voice section detection, and FIG. 3 is a waveform diagram of each generated sound. 1... Telephone speech recognition device 4.5.6... Voice section detection section 11... Selection specification section 12... Voice section detection selection section

Claims (1)

[Claims] A speech signal of a plurality of recognition target words prepared for acceptance in advance is received in accordance with a predetermined order designation, a speech section of the speech signal is detected, and the signal of the speech section is compared with a standard pattern signal. In the speech segment detection method of a speech recognition device, each recognition target word is divided into groups based on features common to each recognition target word, including features due to differences in the number of audio blocks for the audio signal of each recognition target word. a voice section detection section for each group for receiving the voice signal and detecting the voice section; and a voice section detection and selection section for selecting and distributing the input voice signal to each of the voice section detection sections, A speech segment detection method for a speech recognition device, wherein the speech segment detection and selection unit is selectively designated based on the scheduled order designation.