JPS61292199A - 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 recognition device that performs speech segment detection and recognition processing using noise information.

(Prior Art) Conventionally, this type of speech recognition device measures noise in the environment surrounding the speaker or noise caused by input conditions before speech is uttered, and uses the results of the noise measurement as noise information. is used to detect and recognize speech sections. FIG. 5 is a block diagram showing the configuration of a conventional speech recognition device. In the figure, 1 is a voice analysis unit that analyzes input data, 2 is a switch that switches the input source of analysis data from voice analysis unit 1, and 3 is a unit that creates noise information from the analysis data that is input via switch 2. 4 is a noise information storage unit that stores the noise information created by the noise information creation unit 3; 5 is a switch 2;
An analysis data storage unit 6 stores the analysis data inputted through the noise information storage unit 4, and sets a level threshold value using the noise information stored in the noise information storage unit 4. 7 is a recognition processing section that performs recognition processing of the input speech using the analysis data and the speech section information from the speech section detection section 6, and 8 is a controller that controls the entire device. be.

Next, the operation will be explained. First, noise measurement is performed before the first voice is uttered. The controller 8 switches the switch 2 to the noise information creation section 3 side and starts the voice analysis section 1 and the noise information creation section 3. As a result, the input data is input to the audio analysis section 1, where an analysis process for extracting power information and frequency information of the input data is performed at regular intervals (for each frame), and based on the analysis results, the noise information creation section 3, noise information is created and stored in the noise information storage section 4. When the noise measurement is completed, the controller 8 stops the speech analysis section 1 and the noise information creation section 3, and then performs speech section detection and recognition operations. In this case, the controller 8 switches the switch 2 to the voice section detection section 6 side and starts the voice analysis section 1 and the voice section detection section 6. As a result, the input data (f) is analyzed by the voice analysis section l in the same manner as the noise measurement described above, and the analysis result is inputted as voice analysis data to the analysis data storage section 5 and the voice section detection section 6.The voice section detection section 6 detects the sound, A level threshold for detecting the input voice level is set using the voice analysis result and the noise information stored in the noise information storage unit 4 and obtained by noise measurement in advance, and a level threshold for detecting the input voice level is set. The first frame of the section is the starting point and the last frame is the ending point, and the section from this starting point to the ending point is determined to be the speech section.Next, the controller 8 activates the recognition processing section 7.The recognition processing section 7 receives the speech analysis data. Based on the speech section information and noise information outputted from the speech section detection section 5, recognition processing by generally known pattern matching is performed, and the recognition result is output. This figure shows the relationship between noise measurement and voice section detection.

(Problems to be Solved by the Invention) However, in the device with the above configuration, when the voice is input through a telephone or a microphone with weak directivity, there are changes in the surrounding noise of the speaker, and the transmission from the input to the voice recognition device is difficult. If non-pulse noise generated by noise mixed in during noise measurement is accidentally picked up during noise measurement, the noise information will be unique and it will be difficult to compare the state of the noise in the vicinity of the actual voice and the noise obtained by the noise measurement. There was a problem in that the noise information differed greatly, making it impossible to extract accurate speech sections and causing erroneous recognition.

SUMMARY OF THE INVENTION An object of the present invention is to eliminate the above-mentioned problem of erroneous voice section detection due to unique noise information, and to provide an apparatus that cuts out a more accurate voice section and has an excellent recognition rate.

(Means for Solving the Problems) In order to solve the above problems, the present invention provides an analysis means for speech-analyzing input data, an analysis result storage means for storing the analysis results of the analysis means, and an analysis result storage means for storing the analysis results of the analysis means. noise information creation means for creating noise information by measuring noise in a measurement section between an arbitrary start frame and an arbitrary end frame of analysis results stored in a storage means; and noise information from the noise information creation means. noise information storage means for storing;
Detection means for setting a level threshold value using the noise information from the noise information storage means and detecting a speech interval by comparing the content of the analysis result storage means and the level threshold value at the timing when the level threshold value is set; , a recognition means for performing speech recognition based on the output signals of the analysis result storage means and the detection means; and a recognition means for performing speech recognition based on the output signals of the analysis result storage means and the detection means; This is a speech recognition device comprising a control means for updating noise information by increasing the noise information by a fixed amount and detecting a speech section based on the updated noise information.

Preferably, the control means controls each of the means so that the latest contents of the noise information storage means are used as the noise information of the first noise measurement section during the next speech recognition.

(Function) According to the present invention, since the speech recognition device is configured as described above, the technical means functions as follows. The voice input data including the no-voice (noise measurement) section is voice-analyzed by the analysis means and stored in the analysis result storage means. The control means instructs the noise information creation means to create noise information by instructing the first noise measurement section between the first frame of the analysis data and an arbitrary end frame as the start frame. Next, the detection means sets a level threshold value using the noise information inputted through the noise information storage means, and at this timing (end frame 10 level threshold setting time), compares the level of the analysis data with the level threshold value. Then, the audio section from the beginning to the end of the input audio is detected. When the detection means does not detect the start of the voice section, the control means instructs the noise information creation means to update the noise information by instructing the noise measurement section to increase the start frame and end frame by a predetermined amount until the start is detected. . The detection means uses the updated noise information to set a level threshold, and at this timing (
Similarly, the voice section is detected at the update end frame 10 (setting time of the update level threshold). The recognition means recognizes the input speech using the speech segment information obtained in this manner.

Therefore, since the voice section detection is performed using the latest noise information until the voice section is detected, the problems of the prior art described above can be solved.

(Embodiment) FIG. 1 is a block diagram showing a first embodiment of the present invention. In this figure, the same reference numerals as in FIG. 5 indicate the same components. 30 is the noise information creation section 3 in FIG.
A noise information creation unit corresponding to the noise information calculation unit 31
, a start frame setting section 32, and a stop frame setting section 33. The noise information calculation unit 31 receives the voice analysis data (including silent sections) stored in the analysis data storage unit 5 and creates noise information. Start frame setting section 32 and stop frame setting section 33
stores the start frame number and end frame number of noise measurement according to instructions from the controller 8, and provides them to the noise information calculation unit 31.

Next, an example of operation will be explained. First, before audio is input, noise measurement is performed as a primary noise measurement as in the conventional method. All of the voice analysis data analyzed from the voice analysis section 1 in the same manner as before is stored once in the analysis data storage section 5, and the controller 8 stores the start frame number and end frame number predetermined as the section for primary noise measurement. Start frame setting section 32 and stop frame setting section 3
Set each to 3. Next, the noise information creation unit 3o is activated, and the noise information calculation unit 3! reads the audio analysis data of the section corresponding to the frame number set in the start frame setting section 32 and the stop frame setting section 33 from the analysis data storage section 5, creates noise information as in the conventional case, and stores it in the noise information storage section. Store in 4. As a result, the primary noise measurement ends.The zero-order controller 8 activates the voice section detection section 6 and starts voice section detection. The speech section detection unit 6 performs speech section detection as in the conventional method using noise information obtained from the -order noise measurement results,
If the start of the voice section is not detected, start further) y
v A ta constant m (32), the start frame setting section adds a certain number of frames to the start frame, stop frame, and each frame number preset in the stop frame setting section (33). (32), voice analysis data of the section corresponding to the frame number reset in the stop frame setting unit (33) by activating the zero-order noise information creation unit 30 and performing the above-mentioned -order noise measurement. is read from the analysis data storage unit 5 to create noise information and stored in the noise information storage unit 4, and the voice interval detection unit 6 performs zero or more operations to detect the start of the voice interval until the start of the voice interval is detected. By repeating this, the noise measurement section is moved and the latest noise information is provided to the voice section detection section 6. Thereby, the voice section detecting section 6 detects a voice section based on the latest noise information stored in the noise information storage section 4 in the same manner as in the past.

Figure 2 shows the relationship between noise measurement for input data and voice section detection. The section ■ represents the timing for detecting a speech section using the noise information obtained in the section [nl, and the section of the -th noise measurement is the first noise measurement section [11]. Correspondingly, the nth noise measurement interval [
nl indicates that the beginning of voice section ■ has been detected.

Note that the operation after detecting the voice section is the same as the conventional method, so its explanation will be omitted. In the first embodiment, the case where the -th order noise measurement and the period for detecting the speech section are continuous was explained, but even if the period between them is discontinuous, the speech analysis data is treated as if it were continuous. Noise measurement may be performed in the same way as in the case of .

Next, a second embodiment will be described. FIG. 3 is a block diagram showing a second embodiment of the present invention, in which, in addition to the first embodiment, an average noise memory stores the average value of speech analysis data of the section in which noise information used for previous recognition was measured. A portion 41 is provided.

Next, an example of operation will be explained. First, the recognition operation for the -th input data is performed by performing primary noise measurement in the same manner as in the first embodiment, and repeating the noise measurement sequentially until the start of the voice section is determined, and finally the voice section is detected. After performing the recognition process, the noise information used for detecting the speech section is measured, and the average value of the speech analysis data of the section is determined and stored in the average noise storage section 41. Furthermore, for the second and subsequent recognitions of input data, primary noise measurement is not performed, and the average value of the voice analysis data stored in the average noise storage section 41 is used in the analysis data storage section 5 according to a command from the controller 8.
The primary noise measurement is stored in the analysis data storage area for primary noise measurement corresponding to each frame of the voice interval detection interval within the frame, and the primary noise measurement is treated as completed. Give et al.

Then, the noise measurement is repeated for the section designated by the start frame setting section 32 and the stop frame setting section 33 until the start of the voice section is detected, and thereafter the same operation as in the first embodiment is performed. The other operations are the same as in the first embodiment, and furthermore, after the recognition process, the average value of the voice analysis data of the interval in which the noise information used for the voice interval detection and recognition process was finally measured is stored in the average noise storage unit 41. Store in.

Similar to FIG. 1 in the first embodiment, FIG. 4 shows the relationship between noise measurement for input data and voice section detection. In the section [2], the obtained noise information is
. . . represents the timing of detecting a speech section using the noise information obtained in the section ■, and the section for storing the average value of the speech analysis data corresponds to the first noise measurement section [11, and the nth This indicates that the start of the voice section (2) has been detected by the noise measurement section In'l.

In addition, in the second embodiment, the average value of the speech analysis of the section in which the noise information finally used for speech section detection and recognition processing was measured is stored in the average noise storage unit 41 in the device and used for the next recognition. However, the average value of the voice analysis data may be outputted to the outside of the device and received from the outside in the next recognition.

As described above, according to this embodiment, noise information is repeatedly measured until the start of a speech section is detected, and the speech section is detected using the average noise information. Even if non-pulse noise generated by noise mixed in between input and transmission to the device is picked up by noise measurement, a more accurate speech section can be extracted, and a device with an excellent recognition rate can be provided. Furthermore, as in the second embodiment, by not performing primary noise measurement in the second and subsequent recognitions of input data, there is no time taken for noise measurement during continuous recognition, and the speaker's Allows vocalization to occur naturally.

(Effects of the Invention) As described above, according to the present invention, it is possible to accurately cut out a speech section, thereby providing a speech recognition device with an excellent recognition rate.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing a first embodiment of the speech recognition device of the present invention, FIG. 2 is a diagram explaining the operation of the first embodiment,
FIG. 3 is a block diagram showing a second embodiment of the speech recognition device of the present invention, FIG. 4 is a diagram explaining the operation of the second embodiment,
FIG. 5 is a block diagram showing a conventional speech recognition device, and FIG. 6 is a diagram illustrating a conventional speech section detection operation. DESCRIPTION OF SYMBOLS 1... Speech analysis section, 4... Noise information storage section, 5... Analysis data storage section, 6... Speech section detection section, 7... Recognition processing section, 8... Controller, 30 . . . Noise information creation section, 31 . . . Noise information calculation section, 32 . . . Start frame setting section. 33... Stop frame setting section, 41... Average noise storage section.

Claims (2)

[Claims] (1) (a) Analyzing means for speech-analyzing input data; (b)
) analysis result storage means for storing the analysis results of the analysis means; (c) measuring noise in a measurement interval between an arbitrary start frame and an arbitrary end frame of the analysis results stored in the analysis result storage means; (d) noise information storage means for storing the noise information from the noise information creation means; (e) setting a level threshold using the noise information from the noise information storage means; and a detection means for detecting a voice section by comparing the contents of the analysis result storage means and the level threshold at the set timing; (f) based on the output signals of the analysis result storage means and the detection means; (g) updating the noise information by increasing the start frame and the end frame in the noise information generating means by a predetermined amount until the detecting means detects the beginning of the voice section; A speech recognition device comprising: a control means for detecting a speech section based on the noise information obtained. (2) The control means controls each of the means so that the latest content of the noise information storage means is used as the noise information of the first noise measurement section during the next speech recognition. The speech recognition device according to item 1.