JPS61156100A - 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] [Industrial application field] The present invention relates to a speech recognition device.

[Prior Art] Conventionally, in this type of speech recognition device, detection of a speech section is performed based on one type of predetermined threshold value, and if the recognition result is rejected or misrecognized, one speaker is A method was often used in which the same recognition process was repeated by making the user speak again and input the voice again.

[Problems to be Solved by the Invention] In the conventional speech recognition device described above, detection of speech sections is performed based on one type of predetermined threshold value.
If a speech segment is misrecognized or dropped due to improper speech segment detection, the same speech segment detection error will occur even if the speaker speaks again and the speech is re-entered. There is a large risk of repeating misrecognition or misrepresentation, and since the speaker is forced to repeat the same content, there is a problem that the burden on the speaker increases and the quality of service deteriorates.

[Means for Solving the Problems] The present invention stores speech power and characteristic patterns, which are the analysis results of input speech, in a buffer memory, and prepares several types of threshold values for use in detecting speech sections. Then, pattern matching is performed between the characteristic pattern of the detected speech section and the standard pattern to find the degree of similarity.If the recognition result is rejected, the threshold for speech section detection is changed. , read out the power of the input voice stored in the buffer memory, detect the voice section again, read out the feature pattern of the voice section found again from the buffer memory, perform pattern matching, and judge the recognition result. This makes it possible to detect speech sections using multiple thresholds for input speech of a single utterance, increasing the probability of detecting a more appropriate speech section and reducing false detection of speech sections. The probability of rejection or erroneous recognition due to this decreases, and the above-mentioned problems of the conventional device can be solved.

[Example] Embodiments of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram of an embodiment of a speech recognition device according to the present invention.

The speech recognition device of this embodiment includes a speech analysis means 2, an analysis result buffer memory 3, a speech section detection means 4, a standard pattern memory 5. It is composed of a pattern matching means 6 and a recognition result determining means 7.

The speech analysis means 2 inputs the input speech 1, and stores the analysis results of the characteristic pattern 8 and the speech power 9 in the analysis result buffer memory 3. The voice section detecting means 4 reads out the voice power 11 stored from the analysis result buffer memory 3, detects the voice section, and outputs the start and end information 12 of the voice section to the pattern matching means 6. The pattern matching means 6 detects the beginning and end of the voice section.
According to the termination information 12, the characteristic pattern 10 of the voice section stored in the analysis result buffer/memory 3 is read out, and the standard pattern memory 5 (in which a plurality of standard patterns are stored for each word/sentence to be recognized) is read out. The standard pattern 13 is read out from the standard pattern 13, pattern matching is performed between the characteristic pattern of the voice section and the standard pattern, and the similarity 14 is output to the recognition result determining means 7. The recognition result determination means 7 determines the recognition result according to the input similarity 14, and if it is not a reject, it outputs the recognition result 1B and ends the series of speech recognition processing, but if it is a reject, it outputs the recognition result 1B. without outputting recognition results,
A re-recognition instruction 15 is output to the voice section detection means 4 and the pattern matching means 6. When the voice section detection means 4 receives the re-recognition instruction 15, it changes the threshold for voice section detection, reads out the stored voice power 11 from the analysis result buffer 1 memory 3 again, and generates a new voice section. It is detected and output to the pattern matching means 6. When the pattern matching means 6 receives the re-recognition instruction 15, it reads out the characteristic pattern stored in the analysis result buffer memory 3 again according to the start/end information 12 of the new speech section, and performs the same process as the first recognition process. Standard pattern/
The standard pattern 13 is read out from the memory 5, pattern matching is performed, and a new degree of similarity 14 is output to the recognition result determination means 7. The recognition result determination means 7 determines the input new similarity degree 14, and if it is not a reject, outputs the recognition result 16, ends the re-recognition process, the recognition result becomes a reject, and the maximum number of times the re-recognition process is repeated is 1. If the maximum number of times is 2 or more, the re-recognition process is completed by outputting the recognition result 1B of ``Jet'', and if the maximum number of times is 2 or more, the re-recognition process is repeated. If the result is a reject, a recognition result 16 of reject is output, and the re-recognition process is ended.

FIG. 2 is a waveform (temporal change) diagram of an example of the voice power P. In this embodiment, the voice power P for detecting a voice section is
Two threshold values are set, a first threshold value Ptb+ and a second threshold value Ptbz. When the threshold is the first threshold Pth+, the starting point tb where the audio power P matches the threshold Pth+
The starting end duration To starts from +, ends at the end L6 凰 when the audio power P matches the threshold value Pth+ again, and the ending end duration Te starts from this ending te+. Then, a voice section is detected based on the start end duration time Tb and the end end duration time Te. That is, the starting end duration Tb
If it is longer than a certain set time T1, this section is regarded as a voice section, but if the starting end duration Tb does not reach this set time T, it is judged based on the ending end duration Te, and this end end duration Te If it is longer than a certain set time T2, the section of the start end duration Tb is assumed to be not a voice section, and if the end end time Te does not reach the set time T2, this section is added to the previous start end duration Tb. Then, at the next end, it is determined whether it is a voice section or not. *1 threshold P
In the case of ths, the start duration Tb is smaller than the set value T, and the end duration Te is larger than the set value T2, so the section of the start end duration Tb is not regarded as a voice section, and the second threshold value is set. In the case of Pthz, starting point tb2 (starting point t
The start end duration time Tb starts from approximately the same time as b+), and there is a power dip section T- in the middle, but since this section T- is smaller than the set time T2, the start end duration time Tb starts.
b and continues until the next end te2, and the start end duration time T from this start end tI12 to the end end te2
b is larger than the set time T, and this section is regarded as a voice section.

Therefore, even if speech segment detection is first performed using the first threshold value Pth+ and the recognition result is rejected, an appropriate speech segment can be detected by performing speech segment detection using the second threshold value Ptbz. Correct recognition results can be obtained.

Note that in this embodiment, to simplify the explanation, the starting point and the ending point are the point in time when the audio power P starts to exceed the threshold value and the point in time when it starts to go below the threshold value, but a hangover is added to the starting point and the ending point, respectively. In other words, there is also a method in which the starting point is a predetermined time before the voice power begins to exceed the dark value, and the end point is a predetermined time after the voice power begins to fall below the threshold. It is valid.

[Effects of the Invention] As explained below, the present invention stores the power and characteristic patterns that are the analysis results of input speech in a buffer memory, and prepares several types of threshold values for use in detecting speech sections. Then, pattern matching is performed between the characteristic pattern of the detected voice section and the standard pattern to determine the degree of similarity, and the degree of similarity determined is judged and the recognition result is rejected.

Change the voice interval detection threshold, read out the power of the input voice stored in the Hesofa memory, detect the voice interval again, and read out the feature pattern of the voice interval found again from the buffer memory in the same way. matching,
By judging the recognition results, the probability of correctly detecting a voice section for a single utterance of input speech is increased, and the probability of rejection or erroneous recognition due to erroneous detection of a voice section is reduced. This has the effect of reducing the burden on the speaker of re-voicing and enabling more efficient voice input.

[Brief explanation of drawings]

FIG. 1 is a block diagram of an embodiment of a speech recognition device according to the present invention, and FIG. 2 is a diagram showing the relationship between a threshold value for detecting a speech section and a waveform of speech power.・l; Input audio. 2: Voice analysis means. 3: Analysis result buffer fist memory. 4; voice section detection means; 5; standard pattern memory; 6; pattern matching child actor; 7; recognition result determination means; 8; characteristic pattern; 9; voice power. lO; memorized feature pattern. ll: Memorized voice power. 12: Information on the beginning, middle, and end of the audio section. 13: Standard pattern, 14: Similarity, 15: Re-recognition instruction, 16: Recognition result.

Claims (1)

[Scope of Claims] 1. A standard pattern memory in which a plurality of standard patterns are stored in advance for each word/sentence to be recognized; and a voice analysis means for analyzing the voice power and characteristic pattern of input voice; An analysis result buffer memory that stores the voice power and characteristic pattern that are the analysis results of this voice analysis means; and the voice power stored in the analysis result buffer memory is read out, and a voice section is detected from this voice power. a voice interval detection means; and according to the information of the voice interval detected by the voice interval detection means, reading out a characteristic pattern of the voice interval stored in the analysis result buffer memory and reading out a standard pattern from the standard pattern memory. , a pattern matching means for determining the degree of similarity by performing pattern matching between the two, and a recognition result determination means for determining the recognition result according to the degree of similarity, and when the recognition result is not reject, the recognition result is determined. If the recognition result is a reject, the recognition result is not output, and the threshold value used for determining the speech interval in the speech interval detection means is changed, and the recognition process is terminated. A speech recognition apparatus characterized in that a re-recognition process is performed in which the above-described process by the pattern matching means is repeated to obtain a new degree of similarity, and a new recognition result is determined and output according to the degree of similarity. 2. A maximum number of times of the re-recognition process is set, and if the recognition result of the re-recognition process is a reject even if the re-recognition process is repeated the maximum number of times, the recognition result is output as a reject. The speech recognition device according to item 1.