JPH04178699A - Method and device for speech recognition - Google Patents

Abstract

PURPOSE:To prevent the superposition of an abnormal standard pattern for recognition on a normal standard pattern at the time of recording the standard pattern by providing a characteristic extracting section, a standard pattern forming section, a standard pattern memory section for recognition, a standard pattern memory section for synthesis, and a voice synthesizing section. CONSTITUTION:Voice inputs are subjected to the extraction of the characteristic quantity for recognition and the characteristic quantity for synthesis in the characteristic extracting section 2, by which voice sections are determined. These sections are sent to the standard pattern forming section 3 by which the standard patterns for recognition and the standard patterns for synthesis are formed. These patterns are stored in the respective memory sections 4, 5. A speaker utters voices three times and the standard patterns for synthesis formed by the uttered voices are synthesized in the speech synthesizing section 7 and are outputted only when the frame length, the number of the silent sections, etc., of only the voices uttered once among these voices are larger than the predetermined values in the difference from the voices uttered in the other two times. The speaker is made to judge whether these patterns are correct or not and the standard patterns are registered.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a speech recognition method and apparatus, and more particularly to a specific speaker speech recognition method and apparatus for registering a more accurate standard pattern.

Some conventional speech recognition devices register a predetermined number of utterances out of three or more utterances as a standard pattern for synthesis when registering a standard pattern. In addition, there are systems that output a synthesized sound every time a registration voice is uttered, and the speaker determines whether the utterance is correct.

However, when registering the standard pattern, out of 3 or more utterances,
In a speech recognition device that registers a predetermined number of utterances, for example, the second utterance, as a standard pattern for synthesis,
Since it is not possible to confirm the first and third utterances,
When overlapping three audio patterns to create a standard pattern for recognition, even if it is the first or third audio pattern or a defective audio pattern, it will be used and it will not be possible to create a normal standard pattern. .

In addition, with a voice recognition device that outputs a synthesized sound every time a voice is uttered during registration, each time the speaker utters a voice, it must be determined whether the utterance has been input correctly, which places a burden on the speaker. It was big.

■--way.

The present invention was made in view of the above-mentioned circumstances, and
In particular, in speech recognition of a specific speaker, when registering standard patterns, it is possible to prevent an abnormal recognition standard pattern from being superimposed on a normal standard pattern due to a speech interval detection error, etc. This was done with the aim of eliminating the burden of determining whether something is normal or not.

Kazuhiro Toyo In order to achieve the above object, the present invention (1) extracts feature quantities such as spectral information from speech input from a microphone, and creates a recognition standard pattern and a synthesis standard pattern from the extracted feature quantities; and create.

The created standard patterns for recognition are stored, and the created standard patterns for synthesis are stored in correspondence with each standard pattern for recognition, and the input patterns created from the recognition features extracted at the time of voice input are used in advance. The pattern is compared with the stored recognition standard pattern to recognize which standard pattern the input voice corresponds to, and as a result of recognition, a synthesis standard pattern corresponding to the first candidate recognition standard pattern is created. is read from the storage unit of the standard pattern for synthesis and output as a synthesized sound, and the frame length of each utterance is set at the time of registering the standard pattern. , the number of silent intervals, etc. are measured, and the frame length in which only one utterance out of three is measured,
Only when there is a large difference in the number of silent intervals, etc. from the other two utterances and the difference is larger than a predetermined value, a standard pattern for synthesis created by that one utterance is synthesized into speech and output; (2) A feature extraction unit that extracts feature quantities such as spectral information from the voice input from the microphone, and feature extraction. a standard pattern creation section that creates a standard pattern for recognition and a standard pattern for synthesis from the feature values obtained in the section; a standard pattern storage section for recognition that stores the created standard recognition pattern; and a standard pattern storage section that stores the created standard pattern for synthesis. a standard pattern storage unit for synthesis that stores standard patterns corresponding to each standard pattern for recognition; input patterns created from the feature quantities for imperial judgment extracted by the feature extraction unit during voice input; and the above-mentioned pre-stored patterns. A pattern matching unit recognizes which standard pattern the input voice corresponds to by comparing the pattern with the standard recognition pattern in the recognition standard pattern storage unit, and a pattern matching unit recognizes which standard pattern the input voice corresponds to, and as a result of the recognition, the first candidate recognition standard pattern is selected. A speech synthesis unit that reads a corresponding standard pattern for synthesis from the standard pattern storage unit for synthesis and outputs it as a synthesized sound, and performs speech recognition for a specific speaker three times or more per word when registering the standard pattern. When registering standard patterns, the device measures the frame length of each utterance, the number of silent intervals, etc.
Only when the measured frame length, number of silent intervals, etc. of only one utterance out of the two utterances is significantly different from the other two utterances, and the difference is larger than a predetermined value, it is created based on that one utterance. This system is characterized in that the synthesized standard pattern for synthesis is synthesized and outputted by a speech synthesis section, and the speaker judges whether the pattern is correct or not, and then registers the standard pattern.

Hereinafter, the present invention will be explained based on examples.

FIG. 1 is a block diagram for explaining one embodiment of the present invention. As shown in the figure, a feature extraction unit 2 extracts feature quantities such as spectral information from audio input from a microphone 1.
and a standard pattern creation unit 3 that creates a recognition standard pattern and a synthesis standard pattern from the feature amounts obtained by the feature extraction unit 2.
, a recognition standard pattern storage unit 4 that stores created recognition standard patterns, a synthesis standard pattern storage unit 5 that stores created synthesis standard patterns corresponding to each recognition standard pattern, and a voice At the time of input, pattern matching is performed between the input pattern created from the recognition feature extracted by the feature extraction unit 2 and the recognition standard pattern stored in the recognition standard pattern storage unit stored in advance, and it is determined which standard pattern the input voice corresponds to. As a result of the recognition, a pattern matching unit 6 reads out a synthesis standard pattern corresponding to the first candidate recognition standard pattern from the synthesis standard pattern storage unit and outputs it as a synthesized sound. a display section 8 such as a speaker that outputs recognition results, a keyboard 9 equipped with keys for the speaker to perform registration and recognition, and a control section 10 that controls each of the blocks. , a display section 11 for giving operation instructions to an operator (speaker), and the like.

Next, with the above configuration, when registering a standard pattern, 3
The operation of the present invention when registering a standard pattern will be explained using an example in which registration is performed by uttering each time. When registering a standard pattern, the speaker first presses a key for registration on the keyboard 9 and specifies the word number to be registered.

When the control unit 10 detects that the registration key on the keyboard 9 has been pressed, it switches the switch 12, which switches between performing a registration operation and a recognition operation, to the registration side. The speaker inputs the first utterance of the word into the microphone 1, and the input voice is subjected to a feature extraction unit 2, which extracts a feature amount for recognition and a feature amount for synthesis, and determines a speech section. The determined feature amount within the voice section is sent to the standard pattern creation section 3, where a recognition standard pattern and a synthesis standard pattern are created. When this process is finished, the standard pattern creation unit 3 sends a process end signal to the control unit 1o, and after confirming that this operation is finished, the control unit 10 displays the second utterance on the display unit 11 to the speaker. Outputs a prompting message.

The speaker then inputs the second utterance into microphone l,
Similar to the first utterance, the feature extraction unit 2 extracts features for recognition and features for synthesis from the input speech, and determines a speech section. The determined feature amount within the voice section is sent to the standard pattern creation section 3. The standard pattern creation unit 3 creates a standard pattern for recognition from the features for recognition, and sets the frame length and number of silent sections of the standard pattern from the first utterance to 2.
The frame length and number of silent sections of the standard pattern of the first utterance are compared, and the difference therebetween is compared with predetermined thresholds x, y, . . . . As a result, if the difference is smaller than the threshold, a new recognition standard pattern is created by overlapping the first recognition standard pattern and the second recognition standard pattern. Only the one created by the second utterance is retained, and a standard pattern for synthesis is not created from the synthesis feature amount by the second utterance.

- On the other hand, if the difference is larger than a predetermined threshold, the first recognition standard pattern and the second recognition standard pattern are not superimposed and are stored as they are.

Regarding the standard pattern for synthesis, create the standard pattern for synthesis from the feature values for synthesis from the second utterance, and similarly store both the standard pattern for synthesis for the first time and the standard pattern for synthesis for the second time as they are. put. When the processing up to this point is completed, the standard pattern creation section 3 sends a processing end signal to the control section 10, and the control section 10 causes the display section 11 to output a message urging the speaker to speak for the third time. The speaker then inputs the utterance ['' three times into the microphone 1, and similarly to the first utterance, the feature extraction unit 2 extracts features for recognition and features for synthesis from the input voice. , a voice section is determined. The determined feature amount within the voice section is sent to the standard pattern creation section 3. The standard pattern creation unit 3 creates a recognition standard pattern from the recognition feature amount, looks at the number of recognition standard patterns created so far, and if there is one, calculates the number of frames of the recognition standard pattern. The number of silent sections, etc. is compared with the number of frames, the number of silent sections, etc. of the standard pattern for recognition by the third utterance. The difference and respective predetermined thresholds x, y
・If the difference is smaller than the threshold, a new standard recognition pattern is created by overlapping the previously created standard recognition pattern and the standard recognition pattern created by the third utterance. However, as for the standard pattern for synthesis, only the one created by the first utterance is retained, and the standard pattern for synthesis is not created from the feature for synthesis by the third utterance. On the other hand, if the difference is larger than a predetermined threshold, the standard recognition pattern created up to that point and the standard recognition pattern created by the third utterance are not superimposed and both are stored as they are. I'll keep it. Regarding the standard pattern for synthesis, the standard pattern for synthesis is created from the feature quantity for synthesis resulting from the third utterance, and similarly, both the standard pattern for synthesis of the first time and the standard pattern of the third time are stored as they are. In addition, if the number of recognition standard patterns that have been created so far is two, the number of frames and the number of silent sections of the recognition standard pattern created by the third utterance will be compared with the recognition standard pattern created by the first utterance. Compare the number of frames, number of silent sections, etc. of the standard pattern for recognition based on the second utterance, and select the standard pattern with the smaller difference. The selected recognition standard pattern and the recognition standard pattern produced by the third utterance are superimposed to create a new recognition standard pattern. Regarding the standard pattern for synthesis, the standard pattern for synthesis by the third utterance is not created, and the standard pattern for synthesis up to that point is left as is. When the processing up to this point is completed, the standard pattern creation section 3 sends a processing end signal to the control section 10.
When 0 receives this signal, it looks at the number of recognition standard patterns in the standard pattern creation section 3, and if there is one, it stores the recognition standard pattern in the recognition standard pattern storage section 4 and the synthesis standard pattern. Each word is stored in the synthesis standard pattern storage unit 5 together with the word number, and the standard pattern registration process for that word is completed.

On the other hand, the number of standard patterns for recognition in the standard pattern creation section 3 is 2.
In this case, it is detected how many times the recognition standard pattern that is not superimposed was created by utterance, and the synthesis standard pattern created by that utterance is sent to the speech synthesis section 7 to produce a synthesized sound. is output from the speaker 8 as . In this standard pattern, the frame length and number of silent sections differ only once out of three utterances. Therefore, accurate speech section checking may not be performed, and extra noise may be added or the beginning or end of a word may be missing. Therefore, a synthesized sound is output in order for the speaker to judge whether the pattern is standard or abnormal. If the speaker listens to this synthesized sound and determines that this standard pattern is abnormal, he or she presses the cancel key on the keyboard 9, otherwise does nothing. After the control unit 10 receives a signal from the speech synthesis unit 7 indicating that the output of the synthesized sound has ended, the control unit 10 monitors for a certain period of time whether or not the cancel key on the keyboard 9 has been pressed, and if it has been pressed, the control unit 10 , causes the standard pattern creation unit 3 to delete the abnormal recognition standard pattern and synthesis standard pattern. Next, a message prompting the speaker to make another utterance is output on the display section 11, and the speaker inputs the same utterance again. On the other hand, if the cancel key is not pressed for a certain period of time, the other two utterances are considered to be abnormal, and the standard pattern for recognition superimposed on the standard pattern creation section 3 and the standard pattern for synthesis corresponding to it are A message prompting the speaker to make the utterance twice more is output on the display section 11, and the speaker inputs the same utterance twice more. When the recognition standard patterns are superimposed three times, the recognition standard patterns are stored in the recognition standard pattern storage section 4, and the synthesis standard patterns are stored in the synthesis standard pattern storage section 5 together with their word numbers. This is a one-word explanation of the standard pattern registration process. The speaker repeats the action as many times as the number of words he or she wishes to register.

As is clear from the above explanation, according to the present invention, in speech recognition of a specific speaker, when registering a standard pattern, an abnormal recognition standard pattern due to a speech segment detection error etc. can be superimposed on a normal standard pattern. can be prevented, and furthermore,
This eliminates the burden of determining whether the speaker's voice is normal every time he or she utters it.

[Brief explanation of the drawing]

FIG. 1 is a configuration diagram for explaining one embodiment of the present invention. 1. Microphone, 2. Feature extraction section, 3.. Standard pattern creation section, 4.. Standard pattern storage section for recognition, 5..
Standard pattern storage unit for synthesis, 6/pattern matching unit, 7...
Speech synthesis unit, 8...Speaker, 9...Keyboard, 1
0: Control unit, 11: Display unit.

Claims (1)

[Claims] 1. A recognition standard created by extracting feature quantities such as spectral information from audio input from a microphone, and creating a recognition standard pattern and a synthesis standard pattern from the extracted feature quantities. In addition to storing the patterns, the created standard patterns for synthesis are stored in correspondence with each standard pattern for recognition, and the input patterns created from the recognition features extracted at the time of voice input and the previously stored standard patterns for recognition are stored. The pattern is compared with the standard pattern to recognize which standard pattern the input voice corresponds to, and as a result of the recognition, the synthesis standard pattern corresponding to the first candidate recognition standard pattern is used as the synthesis standard pattern. In a specific speaker speech recognition method in which the speech is read from the memory of the speaker and output as a synthesized sound, and utterances are uttered three or more times per word when registering a standard pattern, the frame length of each utterance, the number of silent intervals, etc. are set when registering the standard pattern. If the measured frame length, number of silent intervals, etc. of only one utterance among the three utterances are significantly different from the other two utterances, and the difference is larger than a predetermined value, only that one utterance is measured. A speech recognition method characterized in that a standard pattern for synthesis created by uttering is synthesized and output, and the speaker judges whether the pattern is correct or not and registers the standard pattern. 2. A feature extraction unit that extracts feature quantities such as spectral information from the voice input from the microphone, and a standard pattern creation unit that creates a standard pattern for recognition and a standard pattern for synthesis from the feature quantities obtained by the feature extraction unit. , a recognition standard pattern storage unit that stores the created recognition standard patterns, a synthesis standard pattern storage unit that stores the created synthesis standard patterns corresponding to each recognition standard pattern, and a synthesis standard pattern storage unit that stores the created standard recognition patterns. The input pattern created from the recognition feature extracted by the feature extracting section is compared with the recognition standard pattern stored in the recognition standard pattern storage section stored in advance to determine which standard pattern the input voice corresponds to. a pattern matching unit that recognizes whether the first candidate recognition standard pattern In a specific speaker speech recognition device that utters utterances three or more times per word when registering a standard pattern, the frame length of each utterance, the number of silent intervals, etc. are measured when registering a standard pattern, and one out of three utterances is measured. The frame length measured only by the vocalization of
Only when there is a large difference in the number of silent intervals etc. from the other two utterances and the difference is larger than a predetermined value, the speech synthesis unit synthesizes the standard pattern for synthesis created by that one utterance. A speech recognition device characterized in that a standard pattern is registered by outputting a standard pattern and having a speaker judge whether the pattern is correct.