JPS62267800A - Voice recognition control 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 Industrial Application] The present invention relates to a speech recognition device, and in particular, to a method for reliably and efficiently correcting erroneous recognition or collisions that occur when ambient noise is loud. This invention relates to a voice recognition control method for improving recognition rate.

[Conventional technology]

The speech recognition device calculates a numerical value indicating the degree of similarity between the patterns, that is, distance, by comparing the registered speech pattern and the input speech pattern using a pattern matching method, and calculates the distance, which is a fixed value D set in advance. If it is smaller than □, the candidate with the smallest distance is output as the recognition result, and if it is larger, it is judged that it is questionable to output the recognition result and is rejected. However, the recognition rate may be 10% due to various fluctuation factors in the input voice, such as changes in the speaker's physical condition, ambient noise, switches, etc.
It is difficult to obtain 0% with current technology. Therefore, correction processing for misrecognition etc. is inevitably required.

As described in Japanese Unexamined Patent Publication No. 53-77402, conventional devices display one or more candidates for input speech and allow the speaker to select one.

Also, ¥! fAs described in Japanese Patent Publication No. 59-185394, DTsl is made variable in accordance with the ambient noise level,
Misrecognition and rejections were minimized regardless of the ambient noise level.

[Problem that the invention seeks to solve]

The above-mentioned conventional technology has the following features: (1) Since the button is used to select the item by looking at the display, a worker who cannot take his eyes off the screen, such as a driver driving a car, cannot make corrections.

(2) The candidate display is fixed to one or a plurality of candidates regardless of the ambient noise, etc., and if the ambient noise is a dog, there may be a case where there is no correct answer among the display candidates, making a correction operation impossible.

(3) Even if DTFi is made variable in response to ambient noise, the number of misrecognitions and rejects does not go to zero, and correction processing operations are necessary.

There were other problems.

SUMMARY OF THE INVENTION An object of the present invention is to provide a voice recognition control system that enables safe operation and stable operation regardless of ambient noise (travel noise) by cleaning a voice recognition device used by a driver of an automobile or the like.

[Means for solving problems]

The above object is as follows: (1) It has a means for emitting recognition result candidate speech and a means for audibly selecting one of the candidate speeches, and sequentially emitting the candidate speech while inserting silent intervals into terms with a high degree of similarity. , is achieved by selecting the correct candidate using the selection means by having the speaker who has been listening to it utter a voice during the silent period immediately after the utterance of the correct candidate.

(2) This is achieved by providing means for detecting the ambient noise level and making the number of candidate sounds variable in accordance with the noise level.

(3) This is achieved by providing a volume control means for candidate voice emission and making the volume of the candidate voice variable in accordance with the detection level of the ambient noise level detection means.

(4) This is achieved by providing a candidate voice output speed control means and making the output speed of the candidate voice variable in accordance with the detection level of the ambient noise level detection means.

(5) This is achieved by obtaining a seed pattern for recognition from the registered voice, at the same time providing means for recording and reproducing the voice, and emitting the registered voice using the recognition result candidate voice emitting means.

(6) This is achieved by providing a pitch frequency converting means and converting the pitch frequency of the latter part of the candidate uttered voice and making it a question for listening back.

[Effect]

According to the speech recognition control method of the present invention, (1) The speech recognition control method operates to confirm the recognition result of the input speech and select the correct candidate using speech. Therefore, the operator can perform the verification process without taking his eyes off or using his hands.

(2) When the ambient noise level increases, the number of candidate utterances is automatically increased, that is, the selection range is expanded. As a result, there are many candidate utterances to choose from.
You will not be forced to enter your voice again.

(3) When the ambient noise level becomes dog, the system automatically increases the volume of candidate vocalizations. As a result, the operator can listen to candidate utterances without being disturbed by ambient noise, can make correct selections, and eliminates erroneous operations.

(4) As the ambient noise level increases, the number of utterances required to select candidates increases, but the utterance speed is increased. As a result, the time required to obtain a correct recognition result can be kept constant regardless of the ambient noise level, and operator dissatisfaction can be reduced.

(5) The candidate selection utterance is a registered voice, and the operator listens to and learns the registered voice every time a voice is input.

This allows the voice input for recognition to be matched to the registered voice, reducing misrecognition.

(6) Operate so that the pitch frequency in the latter half of the candidate-selected utterance increases. This allows the operator to easily distinguish it from the input voice, and allows the rise in pitch frequency to signal the beginning of a silent section that cannot be selected, thereby reducing erroneous operations. In addition, since the pitch frequency rise is natural as a response to human voice input, the operator can enjoy the atmosphere of interacting with the device, which can reduce dissatisfaction.

〔Example〕

An embodiment of the present invention will be described below with reference to FIG.

In Fig. 1, 1 is a microphone related to audio input; A 12 is a microphone that adjusts the gain of the input audio signal, performs band limiting and other necessary preprocessing, and then converts it into a digital value.
A D conversion section, 3, a voice section detection section that detects a voice section of the input voice according to a predetermined threshold value from a digital signal, and 4, an analysis of the input voice based on the input digital voice signal, for recognition. A recognition parameter extraction unit extracts the feature parameters to be used; 5 is a standard pattern memory that stores the feature parameter pattern extracted by the recognition parameter extraction unit at the time of registration as a standard pattern at the time of recognition; 6 is a standard stored in the standard pattern memory A standard pattern selection unit selects a pattern, 7 a pattern matching unit that performs pattern matching processing (similarity calculation processing) between the input audio and each standard pattern, and 8 a pattern matching unit that performs pattern matching processing to calculate each similarity for the input audio. A determination section 9 determines the ranking; a microphone B110 relating to ambient noise is an ambient noise detection section detecting the level of ambient noise; and 11 pre-records voice data for informing voice input instructions, operation instructions, rejections, recognition candidate names, etc. 12 is a fixed speech synthesis data selection section that selects fixed speech data; 13 is a speech synthesis section that synthesizes and outputs the speech synthesis data; 14 is a speaker; 15 is each of the above-mentioned sections. 16 is a host device that performs necessary services based on the voice recognition result; and 17 is a console unit that performs necessary manual key operations.

The speech section detection section 3 transmits the start point and end point of the speech to the recognition parameter extraction section 4 and the control section 15. Ambient noise detection unit 1
0 transmits the ambient noise level to the control section 15. The recognition parameter extraction section 4 extracts the input speech and transmits it to the standard pattern memory 5 when registering it as a recognition parameter pattern, and to the pattern matching section 7 during speech recognition processing.

FIGS. 2 and 3 are flowcharts of the voice registration and voice recognition processes, respectively.

First, prior to the voice recognition process, the operator performs a voice ringing operation, and the operation at that time will be explained based on FIGS. 1 and 2. First, from the console unit 17, the control unit 15 and the host device 1 are informed that the registration work is to be performed manually using a key or the like.
6. The control unit 15 instructs each unit to prepare for inputting the registered voice. (Process 30 in Figure 2) Next, the operator inputs the service content code from the console section 17 (Process 31 in the same figure), and then makes an input reminder sound (Process 3 in the same figure).
2), input the voice corresponding to the content code. For example, a code a is inputted from the console section 17, and then the input reminder is uttered followed by the utterance of ``A''.

When the operator inputs voice from the microphone A1 (process 33 in the figure), the AD converter 2 converts the voice into a digital signal, the voice section detector 6 detects the digital voice signal, and the recognition parameter extractor 4 converts the voice into a digital signal. The feature parameters of the voice are extracted (process 34 in the figure) and stored in the standard pattern memory 5 (process 35 in the figure).

By repeating the above process, a standard speech pattern for recognition is registered. For example, content codes b, c, d.

"B" and "Shi" for the key human power of e...

Say "di", "ee"... This completes the voice registration work.

Next, the speech recognition processing operation will be explained based on FIGS. 1 and 3.

First, the console unit 17 informs the control unit 1 that recognition processing is to be performed.
5 and the host device 16. The control unit 15 instructs each unit to prepare for voice input in recognition processing. (Process 40 in FIG. 3) When these preparations are completed, an input reminder message to prompt the operator to input voice is emitted from the speaker 14 via the voice synthesis section 13. This is done by transmitting the corresponding message data from the fixed speech synthesis data memory 11 to the speech synthesis section 13 via the fixed speech synthesis data selection section 12 (process 41).

Accordingly, when the operator inputs f voice from the microphone A1 (same process 42), the AD converter 2 converts the voice into a digital signal, and the voice section detecting unit S detects the digital voice signal and performs recognition. The parameter extraction unit 4 extracts the characteristic parameters of the voice (process 43).

The pattern matching unit 7 performs a pattern matching process between the characteristic parameter pattern of the input voice and each climbing type voice pattern stored in the standard pattern memory 5 and selected by the mark pattern selection unit 6. The degree of similarity of each standard voice pattern to the input voice is calculated and transmitted to the determination unit 8 (processing 4).
4). - The determination unit 8 ranks each standard pattern based on the degree of similarity, and uses the ranking as a result along with the degree of similarity. 1f Inform the IJ department 15. Similarity is the highest (
The most probable) is ranked first, then the second, etc.
... (same process 45).

If the most probable similarity value for the input voice is larger than a predetermined threshold and it is doubtful to output it as a result, the control unit 15 prompts the input again. In order to do so, a message reminding the user to re-enter is emitted from the speaker 14 via the voice synthesis section 13. This transfers the corresponding message data from the fixed speech synthesis data memory 11 to the fixed speech synthesis data selection section 12.
This is done by transmitting the information to the speech synthesis unit 13 (processing 46).

If it is not rejected, the control unit 15 checks the recognition result, and the determination unit 8 determines the recognition result. According to the @ position, the corresponding candidate name speech synthesis data is sent from the speech data of the recognition candidate names stored in advance in the fixed speech synthesis data memory 11 to the speech synthesis section 13 via the fixed speech synthesis data selection section 12. 4 (speaker) 4 (same process 47).

For example, if the first place is a standard pattern corresponding to code a, candidate name speech synthesis data representing recognition candidate name a is synthesized and ``A'' is uttered. The second place is fi! corresponding to code b! If it is a pattern, then candidate name speech synthesis data representing the recognition candidate name is synthesized and a "beep" is emitted, and the same is repeated.

This sound emission is controlled by the control unit 15, and is performed with an appropriate interval (silent interval) between the sound emission and the sound emission. Furthermore, the level at which the sound should be emitted is determined by collecting ambient noise with the microphone B9, monitoring the ambient noise level before and after the input voice is uttered, detected by the ambient noise detecting section 10, with the control section 15, and using a trap. Processing 42\Figure 1A is a diagram showing the cumulative recognition rate. In the figure, (a) is when there is no ambient noise, and (b) is when there is ambient noise. (a) ranks 3rd. The recognition rate for the candidates is 100. This means that if you select the top three recognition result candidates, the correct recognition result will definitely be included.As shown in the figure, ambient noise is mixed in. Overall, the similarity between the registered speech and the input speech is small, and the recognition rate decreases.In (b), the recognition rate is 100chi for the top five places.The degree of decrease in the recognition rate also depends on the ambient noise level. is proportional to.

In this embodiment, based on the above knowledge, the number of candidate name announcements is varied according to the level of the three surrounding sounds.

For example, answer "Hi." This sound is from microphone A
1, the audio is converted into a digital signal by the AD converter 2, detected by the audio section detector 3, and then sent to the controller 1.
5 (same process 48).

If there is any voice input immediately after the candidate name is uttered, that is, during a silent period, the control unit 15 determines that this is a correct recognition result, and transmits the content code to the host device (process 49). The host device then performs a predetermined service based on the results.

A predetermined number of candidate names will be announced based on the ambient noise level.
It is turned back under the control of the control unit 15 (same process 50).
If there is no response from the operator after a predetermined number of vocalizations, it is assumed that the operator has ejected, and a voice prompting for re-input is issued (processing 46).

The control unit 15 monitors the key input from the console unit 17 to end the voice recognition process, and if it is not finished, issues an input reminder again and returns the same voice recognition process (same process 51).
.

In the above explanation, we have described a specific speaker recognition device in which the operator registers his or her own voice, but it is a standard pattern that is created in advance by processing such as averaging from the voice patterns of many people (average pattern). If registered, it will operate as a speaker-independent recognition device.

In this case, the registration work shown in FIG. 2 is not necessary.

Furthermore, although the recognition similarity was determined by pattern matching processing, other methods such as function identification may be used.

Further, the microphone B9 for collecting ambient noise may be replaced by the microphone A1 for audio input, and furthermore, the input to the ambient noise detection section 10 may be the output of the converting section 2.

Further, although the response judgment to the candidate name announcement was based on the presence or absence of voice, this may be replaced with a process that recognizes "yes" and "no". Furthermore, the response may be made by inputting a key to the console section 17, for example, instead of by voice.

Furthermore, although the candidate name announcement was performed by voice synthesis, there is no need to be limited to this; for example, a device capable of recording and reproducing voice, such as a cassette tape recorder, may be used.

Fig. 4 shows another recognition processing flowchart in the embodiment of Fig. 1. In FIG. 4, the same symbols as in FIG. 3 indicate the same processing.

In FIG. 3, the number of recognition candidate names is determined only by the ambient noise level, but in FIG. 4, a threshold value for rejection is further added.

Determining whether a predetermined number of vocalizations have been completed (processing 50)
If the similarity of the candidate to be emitted next is less than or equal to the reject threshold, it will be emitted, but if it is greater than that, the emitted voice will be discontinued even if it does not reach the predetermined number, and it will be considered a reject. Process 52), a re-input reminder is issued.

FIG. 5 shows another embodiment of the present invention. In FIG. 5, the same symbols as in FIG. 1 indicate the same parts.

Reference numeral 18 denotes a volume control unit that controls the amplitude, that is, the volume, of the audio signal synthesized by the audio synthesis unit 13.

The volume control section 18 controls the volume of the audio signal according to the ambient noise level detected by the ambient noise detection section 10, and controls the volume of the audio signal from the speaker 14.
Let out a voice from As a result, when the surrounding plaster becomes large, the volume of the candidate name is increased to make it easier for the operator to hear the sound. Fifth
The registration and recognition processing of the embodiment shown in the figure is the same as that of FIG. 2.3.4.

FIG. 6 shows yet another embodiment of the present invention.

In FIG. 6, the same reference numerals as in FIG. 1 indicate the same parts.

Reference numeral 19 denotes a speech rate control unit for the voice synthesized by the voice synthesis unit 13. The speech rate control section 19 controls the speed of the candidate name voice based on the ambient noise level detected by the ambient noise detection section 10, and outputs the voice from the speaker 14. As a result, the listening time of the operator will not be increased even if the number of candidate names is increased even if the ambient noise is low. Note that the length of the silent section inserted between candidate name announcements may be fixed or may be varied depending on the ambient noise level.

The registration and recognition processing in the embodiment of FIG. 6 is the same as that of FIG. 2.3.4.

FIG. 7 shows another embodiment of the present invention. In FIG. 7, the same symbols as in FIG. 1 indicate the same parts.

Reference numeral 20 denotes a speech synthesis data extraction section that compresses information on input speech based on the input digital speech signal and extracts data used for synthesis, and 21 stores the data extracted by the registered voice synthesis data extraction section 2o. The standard speech synthesis data memory 22 is a standard speech synthesis data selection unit that selects the standard speech synthesis data stored in the standard speech synthesis data memory 21.

In the embodiment shown in FIG. 1, the recognition candidate name is uttered based on the candidate name speech synthesis data stored in the fixed speech synthesis data memory 11 in advance. In this embodiment, a standard speech pattern for recognition is obtained at the time of registration, and at the same time, standard speech synthesis data corresponding to the standard speech pattern for announcing the recognition candidate name is obtained.

Accordingly, the registered voice is used to announce the recognition candidate name during recognition.

FIG. 8 shows a registration processing flowchart of this embodiment. In FIG. 8, the same symbols as in FIG. 2 indicate the same processing. When the operator inputs audio from microphone A1 (same process 3
3), the AD conversion unit 2 converts the input voice into a digital voice signal, the recognition parameter extraction unit 4 extracts the characteristic parameters of the voice (process 34), and stores it in the standard pattern memory 5 (process 35). ). At the same time, the speech synthesis data extraction section 20 compresses the information of the speech, extracts the data (process 36), and stores it in the standard speech synthesis data memory 21 (process 37).

By repeating the above process, a standard speech pattern for recognition is registered, and data of the same speech is stored as standard speech synthesis data.

The recognition process is almost the same as that shown in FIG. 3.4, but the recognition candidate voices in process 47 are different. The recognition candidate speech is registered as a standard speech pattern in the previous registration process, and the same standard speech synthesis data as the speech is sent from the standard speech synthesis data memory 21 to the speech synthesis section 13 via the standard speech synthesis data selection section 22. tell it to
A sound is emitted from the speaker 14. By doing this, the operator listens to the voice he or she has registered as the recognition candidate voice.

FIG. 9 shows another embodiment of the present invention. In FIG. 9, the same reference numerals as in FIG. 7 indicate the same parts.

23 is a pitch frequency conversion unit that converts the pitch frequency of the audio signal synthesized by the audio synthesis unit 13.

The registration process in this embodiment is the same as that in FIG. 8, and the recognition process is almost the same as in the embodiment in FIG. 7, but the recognition candidate voices in process 47 are different. In this embodiment, when a recognition candidate voice is emitted,
The control unit 15 controls the pitch frequency conversion unit 23 to increase the pitch frequency of the latter half of the recognition candidate voice, and listens to the voice as a question.

〔Effect of the invention〕

According to the present invention, (1) the recognition result of the input voice can be checked and the correct candidate can be selected by voice, so the operator can perform other tasks without looking away.

(2) The range of options for correct or incorrect recognition is expanded according to the ambient noise level, which has the effect of eliminating the need to input voice again for rejection.

(3) Since the sound volume of the recognition result candidates increases according to the ambient noise level, the candidate sounds can be heard without being disturbed by the surrounding noise, allowing for correct selection and fewer erroneous operations.

(4) The number of voices emitted from recognition result candidates increases in accordance with the ambient noise level, but at the same time, the speed at which they are emitted also increases, which has the effect of resolving operator dissatisfaction without changing the time it takes to obtain a correct recognition result.

(5) The candidate selection voice is a registered voice, and since the operator listens to and learns the registered voice each time an input is made, this has the effect of reducing erroneous recognition (by matching the input voice to the registered voice).

(6) Since the pitch frequency of the latter half of the candidate voice is raised, the candidate voice and the input voice can be easily distinguished, and the increase in pitch frequency also marks the beginning of a silent section and serves as an accent for questions to be heard. This allows the operator to experience the atmosphere of interacting with the device, which has the effect of reducing dissatisfaction.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention, FIG. 1A is a graph showing the recognition rate for candidate positions, and FIG.
3 is a flowchart of the registration process of the embodiment shown in FIG. 1, FIG. 4 is a flowchart of another recognition process of the embodiment of FIG. 1, and FIG. FIG. 6 is a block diagram showing yet another embodiment of the present invention. FIG. 7 is a block diagram showing another embodiment of the present invention. FIG. 8 is a block diagram showing another embodiment of the present invention. FIG. 9 is a block diagram showing still another embodiment of the present invention. 1...Microphone A 4...Recognition parameter extraction section 5...Standard pattern memory 9...Microphone B 10...Ambient noise detection section 11...Fixed speech synthesis data memory 13...Speech synthesis 14...Speaker 18...Volume control section 19...Speech rate control section 20...Speech synthesis data extraction section 21...Standard speech synthesis data memory 23...Pitch frequency conversion section E-shu , Capture 11 Self Ik (I standing) Atsushi 3 figure Tsuta 4 figure

Claims (1)

[Claims] 1. In a speech recognition device that performs recognition by comparing a speaker's speech input pattern with a standard speech pattern, as a result of the comparison, the standards obtained as recognition result candidates are ranked in descending order of similarity. means for sequentially emitting voice patterns as sounds; means for detecting a sound to be emitted as a speech when a speaker who listens to the sequentially emitted standard sound patterns selects one of the standard sound patterns; and an ambient noise level. means for detecting the noise level; and means for variably controlling the number of standard sound patterns sequentially emitted from the sound emitting means according to the detected noise level;
A voice recognition control method characterized by comprising: 2. In the voice recognition control system according to claim 1, the number of standard voice patterns sequentially emitted from the voice emitting means is determined based on the noise level detected by the noise level detection means and the voice input pattern. A speech recognition control method characterized by variable control according to both a preset threshold for similarity obtained when comparing with a standard speech pattern. 3. The voice recognition control method according to claim 1, wherein the volume of the voice emitted from the voice emitting means is variably controlled in accordance with the noise level detected by the noise level detection means. . 4. In the voice recognition control system as set forth in claim 1, the voice emission speed of the voice emitted from the voice emission means is variably controlled in accordance with the noise level detected by the noise level detection means. Recognition control method.