JPH01310399A - Speech recognition device - Google Patents

Abstract

PURPOSE:To accurately detect a speech section even in environment where an external noise is loud by detecting the distance between the lips of a speaker and a microphone and entering the section wherein the lips move as a next candidate for the speech section. CONSTITUTION:A distance sensor 3 provided to a microphone device 1 detects and outputs variation in the distance between the microphone 2 and the lips of the speaker 21 including the motion of the lips and their peripheral part when the speaker 21 speaks to a distance calculation part 7. The calculation part 6 converts the detection signal into a digital value and outputs it to a voicing section detection part 7. The detection part 7 uses a threshold value which is determined adaptively to regard the section wherein the largest variation in the distance between the microphone 2 and lips is obtained as the voicing section, and outputs it to a time normalization part 8 as the next candidate for a speech section signal outputted from a speech section detection part 5. When the external noise is too loud to detect the speech section through the microphone 2, the voicing section is utilized to find the speech section.

Description

[Detailed description of the invention]

[Object of the Invention] (Industrial Application Field) The present invention relates to a speech recognition device. (Prior art) A speech recognition device inputs the speech generated by a speaker, detects a speech section from the speech signal, and temporally normalizes the signal of this speech section to obtain feature quantities and a standard pattern. This is a device that calculates the degree of similarity 'a between ``a'' and outputs the category with the highest score as a voice recognition result, and is used in various devices that operate automatically by operations using human voice. ing. Therefore, in speech recognition using a speech recognition device, as a preprocessing step, the speech interval of the speech generated by the speaker is
It is necessary to perform processing to detect the In order to detect a speech interval in a conventional speech recognition device, the voice of the speaker is input through a microphone, and an appropriate threshold value is used based on the acoustic feature parameters obtained by analyzing the input voice. A voice section is detected by detecting the start and end of the voice section. (Problem to be Solved by the Invention) However, in the method of detecting speech intervals in such a conventional speech recognition device, in an environment with large external noise, external noise may be heard in addition to the voice generated by the speaker. It may also be input into the microphone. In this case, when analyzing the voice, even a minute amount of the original voice component and other external noise components are superimposed, so it may be difficult to accurately detect the voice section. This is particularly noticeable when the noise level is high or sudden noise occurs. The present invention has been made based on the above-mentioned circumstances, and an object of the present invention is to provide a speech recognition device that can accurately detect speech sections without being affected by external noise. [Structure of the Invention] (Means for Solving the Problem) In order to achieve the above object, the speech recognition device of the present invention includes a distance sensor that detects the distance between the speaker's device and the microphone, and a distance sensor that detects the distance between the speaker's device and the microphone. A distance calculation means for converting the detected signal into a digital quantity, and a stable distance ilitm is extracted from the distance time series of the signal output by this distance calculation means, and the utterance period of the voice is detected from the movement of the lips and the surrounding area. The invention is characterized in that it includes a vocalization section detecting means for detecting a vocalization section, and uses the vocalization section detected by the vocalization section detecting section as the next candidate for the vocalization section. (Function) In other words, the distance sensor detects the distance between the speaker's lips and the microphone when the speaker speaks, and the section in which B is moving is regarded as the section in which the speaker is generating sound, and the speech section is determined. Enter as the next candidate. As a result, even if it is difficult to detect a voice section from a voice signal including loud noise input from a microphone in an environment with large external noise, the voice section can be accurately detected using the utterance section. (Example) Examples of the present invention will be described below with reference to the drawings. An embodiment of the configuration of the speech recognition device of the present invention will be described with reference to FIG. In the figure, 1 is a microphone device in which a microphone 2 and a distance sensor 3 are integrated! As shown in FIG. 2, the speaker 21 wears and uses the device. The microphone 2, which is a voice input section, is of a close-talking type and is provided at a position a certain distance in front of the lips 22 of the speaker 21. Distance sensor 3
is a sensor that measures the distance between the microphone 2 and the lips 22 of the speaker 21. The distance between the microphone 2 and the lips 22 changes as the lips 22 and the surrounding area move when speaking. Detect changes in Therefore, the distance sensor 3 is provided at a position where the distance between the microphone 2 and the lips 22 can be accurately detected. As the distance sensor 3, an infrared sensor, an ultrasonic sensor, or the like is used, and among them, an infrared sensor is preferable because it has less noise. The microphone 2 and the distance sensor 3 are integrally supported by a support member 1a, which is worn by a speaker 21. Note that the microphone 2 and the distance sensor 3 are configured to send signals to the main body of the apparatus via a signal line (not shown). Reference numeral 4 denotes an acoustic analysis section that receives the audio signal input from the microphone 2, extracts its acoustic characteristic parameters, and outputs the signal to the audio section detection section 5. The voice section detection section 5 receives the signal from the acoustic analysis section 4, detects a voice signal, and outputs the signal to the time normalization section 8. 6 is a distance calculation unit which receives the detection signal from the distance sensor 3, converts it into a digital quantity, and converts the signal into a digital quantity.
It is output to section 7. The vocalization section detecting section 7 receives the signal from the distance calculation section 6, detects the vocalization section, and outputs the signal to the time normalization section 8. Reference numeral 8 denotes a time normalization unit which receives the voice interval signal and the utterance interval signal respectively output from the voice interval detection unit 5 and the utterance interval detection unit 7, obtains temporally normalized feature quantities from each signal, and calculates the similarity. It is output to the degree calculation section 9. The similarity calculation section 9 receives the signal from the time normalization section 8 and performs a similarity calculation between it and the standard pattern 10. A case in which speech recognition is performed by the speech recognition device configured in this manner will be described. The voice generated by the speaker 21 wearing the microphone device 1 is input to the microphone 2, and the acoustic analysis unit 4 extracts acoustic characteristic parameters from this voice signal. The signal of the extracted feature quantity is output to the speech interval detection unit 5, and the speech interval detection unit 5 uses a part of the feature quantity (for example, a power sequence) to determine the speech interval according to an adaptively determined threshold value. Detect the start and end of. This voice section signal is output to the time standardization section 8. On the other hand, the distance sensor 3 provided in the microphone device 1 detects a change in the distance between the microphone 2 and the lips 22 due to the movement of the lips 22 and its vicinity when the speaker 21 utters a voice. The signal is output to the distance calculation section 6. The distance calculation section 6 converts this detection signal into a digital quantity and outputs it to the utterance section detection section 7. The utterance section detecting section 7 uses an adaptively determined threshold to detect the section in which the distance variation between the microphone 2 and 822 is the largest, and regards this as the utterance section. This vocalization section signal is output to the time regulation section 8 as the next candidate for the vocalization section signal outputted from the vocalization section detecting section 5. The time regulation unit 8 receives the voice interval signal and the utterance interval signal output from the voice interval detection unit 5 and the utterance interval detection unit 7, respectively, and receives the feature ff1 (for example, bandpass output) extracted by the acoustic analysis unit 4. part) is temporally normalized within the speech interval and within the utterance interval to obtain two types of feature quantities, and output these signals to the similarity calculation unit 9. The similarity calculation unit 9 performs similarity calculation between these two types of feature amounts and the standard pattern 10, and outputs the category showing the highest score as a speech recognition result. Here, a voice interval is determined based on the voice generated by the speaker 21, and a voice interval is determined by focusing on the movement of the lips 22 and its vicinity when the voice is generated, and this generated interval is used as the next candidate for the voice interval. Therefore, if the external noise is large and the external noise is input to microphone 2 in addition to the voice, and it is difficult to detect the vocal interval from a small amount during analysis, the vocal interval can be used. The speech interval can be found by The obtained voice section is accurate and matches the original voice section. Next, the operations of the utterance period detecting section 7 and the voice section detecting section 5 will be explained, and the features of the apparatus of the present invention will be explained. FIG. 2 shows the microphone 2 output from the distance calculation section 6.
2 shows an overview of the internal processing flow of the voice section detecting section 7 that searches for a voice section from the distance between the lips 22 of the speaker 21 and the lips 22 of the speaker 21. First, a smoothing process is performed on the time series data D (n) between the microphone 2 and the lips 22 to extract a time series d (n) of stable distance variation (SPI). Next, calculate the average distance d^ in the interval n-1 to 10 of this time series d (n) that can be considered as a silent interval 5P2
), the threshold value dTll (-dT + do) is determined by adding the adaptively determined offset do to this (
S P 3). The search for vocalization sections SA and SE is performed using d (n) > d T
When the section ll continues for several frames or more, the position where d (n) > d Tll for the first time becomes the starting point,
Furthermore, the interval d (n) > d Tll is continuously 1
After 0 frames or more, d (n) < d Tl
Only when the interval l continues for several frames or more, d
The position where (n) < d Tll is determined as the end, and a signal is sent to the time regulating unit 8 with this SA and SE section as the utterance section (SF3). Furthermore, the voice section detecting section 5 also searches for a voice section using a similar method. In this case, it is determined by the feature amount extracted from the audio signal input from the microphone 2, and the feature amount is not necessarily based on the voice, but may include external noise components around the microphone. FIGS. 3 and 4 are diagrams showing the processing of the vocalization section detection section 7 and the speech section detection section when large external noise is present before and after the speech section. Since the vocalization sections SA and SE detected by the vocalization section detecting section 7 shown in FIG. It is possible to detect the interval in which speech is considered to have been generated based on mouth movements. On the other hand, in the processing of the speech section detection unit 5 shown in FIG. 4, even if the adaptively determined threshold ii p T11 is used, the speech section ST is accurate due to noise components. It becomes impossible to detect ET, and the section S containing noise components
F and EF are erroneously detected. Detection of a voice section is a fatal error for the voice recognition device. In order to avoid such a situation, the utterance section SA is not affected by external noise even in an environment with relatively large external noise.
, EA as the next candidate for the speech section, it is possible to realize a speech recognition device with a more stable and high recognition rate. Note that the present invention is not limited to the embodiments described above, and can be implemented with various modifications without changing the gist. The microphone for voice input is not limited to the normally used close-talking type microphone, but may be changed depending on the device for capturing voice to which the device of the present invention is applied. For example, a telephone receiver can be used as a microphone. In this case, the telephone receiver is fixed at a fixed position, and the distance sensor is integrally attached to the telephone receiver.

【Effect of the invention】

As explained above, according to the speech recognition device of the present invention, the distance sensor measures the change in the distance between the speaker's port and the microphone when the speaker speaks, and the range in which B moves is uttered. By detecting it as a segment and entering it as the next candidate for the speech segment, it is possible to use the speech segment even when it is difficult to detect the speech segment from the voice input from the microphone in a noisy environment. It is possible to accurately detect voice sections without being affected by external noise, and it is possible to obtain a high recognition rate and stable accuracy that is not restricted by the usage environment.

[Brief explanation of the drawing]

FIG. 1 is a system configuration diagram showing an embodiment of the speech recognition device of the present invention, FIG. 2 is an explanatory diagram showing a microphone device in the same embodiment, and FIG. FIG. 4 is a diagram showing the state of voice section detection by the voice section detecting section, and FIG. 5 is a diagram showing the state of voice section detection by the voice section detecting section. DESCRIPTION OF SYMBOLS 1... Microphone device, 2... Microphone, 3... Distance sensor, 4... Acoustic analysis part, 5...
Speech section detection unit, 6... Distance calculation unit, 7... Vocalization section detection unit, 8... Time normalization unit, 9... Similarity calculation unit. Applicant's representative Patent attorney Takehiko Suzue Figure 2 Figure 3 Figure 4 Figure 5

Claims (1)

[Claims] Input the voice generated by the speaker from the microphone,
In an apparatus that performs speech recognition by detecting a speech section from the speech signal and performing a similarity calculation between the feature amount obtained by temporally normalizing the signal of this speech section and a standard pattern, A distance sensor that detects the distance between the lips and the microphone, a distance calculation means that converts the signal detected by this distance sensor into a digital quantity, and a stable distance based on the distance time series of the signal output by this distance calculation means. utterance period detecting means for extracting the amount of utterance and detecting the utterance section of the voice from the movement of the lips and the vicinity thereof, the utterance section detected by the utterance section detecting section being set as a next candidate for the voice section. Characteristic voice recognition device.