JPH0412478B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a monosyllabic speech recognition device for recognizing speech.

Conventionally devised monosyllabic speech recognition devices are configured to segment input speech into consonant parts and vowel parts, and recognize the input speech using the average characteristic pattern of the temporal center of each part. ing.
However, it is generally well known that the patterns of consonant parts are often non-stationary in time, and if the average feature pattern described above is used to identify consonant parts, the non-stationary features of consonants can be Become clear. Therefore, especially for identifying consonant parts, it is desirable to use characteristic patterns for multiple frames within a consonant section. Note that a frame refers to each sampling point in time when input audio data is sampled at fixed time intervals in order to generate a characteristic pattern.

On the other hand, devices that recognize input speech using characteristic patterns of all frames within a speech section have also been devised as monosyllabic speech recognition devices.
Although such a device can compensate for the shortcomings of the above-mentioned devices, the characteristic pattern for vowel parts is stable over a relatively long period of time (for example, about 100+ msec), and the Recognizing input speech using frame feature patterns results in spending more time than necessary, so in order to shorten the time required for recognition, recognition is performed using average feature patterns as described above. This is desirable.

In view of the above-mentioned drawbacks, the present invention first calculates an average feature pattern by averaging feature patterns for a portion that is certain to be a vowel part of a monosyllabic speech, that is, a few frames before the end of a speech section.
Next, feature patterns are found sequentially from the beginning of the word and the distance from the average feature pattern is calculated, and when the distance is smaller than the threshold for several frames, or the distance is smaller than the threshold. When this continues for several frames in a row, the analysis for finding the feature pattern ends, and the input speech is recognized using the feature pattern that has already been found, thereby reducing the time required for recognition and reducing the demands on the hardware. The present invention provides a monosyllabic speech recognition device that can reduce the speed of processing performed.

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

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

In FIG. 1, reference numeral 1 denotes a speech section detection unit which detects time points corresponding to the beginning and end of a monosyllabic speech input a, and outputs beginning and end detection signals. Reference numeral 2 denotes a voice holding unit that holds data of the monosyllabic voice input a from the time when the word beginning detection signal sent from the voice section detecting unit 1 is obtained to the time when the word end detection signal is obtained. Reference numeral 3 denotes an average feature pattern generation unit, which generates a pattern from a point several frames (several sampling points) before, for example, 10 frames before the end of a word detection point in the voice data held in the voice holding unit 2.
Several frames (several sampling points), e.g. 5
Frames of audio data are sampled at regular intervals to obtain feature patterns, which are averaged to generate and output an average feature pattern. Reference numeral 4 denotes a characteristic pattern generating section which samples the voice data held in the voice holding section 2 sequentially from the beginning of the word at regular intervals to generate and output characteristic patterns. Reference numeral 5 denotes a feature pattern holding unit that holds the feature pattern output from the feature pattern generating unit 4. 6 is an inter-feature pattern distance calculation unit that calculates the distance between the average feature pattern sent from the average feature pattern generation unit 3 and the feature pattern sent from the feature pattern generation unit 4, and calculates the calculated distance. Output. Reference numeral 7 denotes a threshold determination section, which compares the distance output from the inter-feature pattern distance calculation section 6 with a predetermined threshold to determine the magnitude, and outputs a count increase signal when the distance is smaller than the threshold. Reference numeral 8 denotes a counting unit, which increases the count value by 1 each time it receives a count increase signal that is the output of the threshold value determination unit, and when the count value reaches a predetermined value, causes the characteristic pattern generation unit 4 to generate a characteristic pattern. At the same time, it outputs a characteristic pattern generation end signal for terminating the process, and at the same time outputs a characteristic pattern output command signal for outputting the characteristic pattern held in the characteristic pattern holding section 5. Reference numeral 9 denotes a speech identification section which identifies input speech using the output of the feature pattern holding section 5 and outputs a monosyllabic speech recognition result b.

The operation of the apparatus configured as described above will be specifically explained below.

First, input speech that has been low-pass filtered with a cutoff frequency of 5 KHz is A/D converted at a sampling frequency of 10 kHz, and the discrete signal is detected by the speech section detection section 1 using, for example, energy level to detect the beginning of a word, and the speech holding section 2 The holding of the discrete signal is started from the time when the beginning of the word is responded to, and the holding of the discrete signal is ended when the voice section detecting section 1 detects the ending of the word. Next, in the average feature pattern generating section 3, for example, the time width is 12.8 m.
While shifting the Hamming window of sec by 6.4 msec (at this time, the frame period is 6.4 msec), it is added to the discrete signal held in the voice holding unit 2, and the hamming window is added to the discrete signal held in the voice holding unit 2 by 70.4 msec from the end of the word (from the end frame of the word).
10 frames) From the previous point, 32 msec from the end of the word
The 14th-order PARCOR coefficients are obtained for the five frames of discrete signals up to the previous point in time, and averaged to generate an average feature pattern, which is output to the inter-feature pattern distance calculation section 6. On the other hand, in the same way as the average feature pattern generating part 3, the characteristic pattern generating part 4 adds the Hamming window, which has a time width of 12.8 msec, to the discrete signal held in the audio holding part 2 while shifting the Hamming window by 6.4 msec, for example. The PARCOR coefficients are sequentially calculated from the beginning of the word at a frame period of 6.4 msec.
, and is output to the feature pattern holding section 5 and the inter-feature pattern distance calculation section 6. The characteristic pattern holding section 5 holds the characteristic pattern generated by the characteristic pattern generating section 4 every 6.4 msec. On the other hand, in the feature pattern distance calculation unit 6, the PAPCOR coefficient as an average feature pattern and the PARCOR coefficient as a feature pattern obtained every 6.4 msec,
For example, the Euclidean distance is calculated and the calculation result is output to the threshold determination section 7. Next, in the threshold determination section 7,
For example, if the threshold is set to 0.2, and the Euclidean distance value becomes smaller than the threshold, the count of the counter 8 is set to 1.
Outputs a count increment signal to increase the count. When the count increase signal is applied, for example, when the count value reaches 5, the counting unit 8 outputs a characteristic pattern generation end signal that causes the characteristic pattern generation unit 4 to terminate characteristic pattern generation, and at the same time outputs the characteristic pattern generation end signal to the characteristic pattern holding unit 5. A feature pattern output command signal is output to the feature pattern holding unit 5 to cause the voice identifying unit 9 to output the feature pattern held in the voice identifying unit 9 . The speech identification section 9 can identify the monosyllabic speech input a using the characteristic pattern obtained in this way, and output the monosyllabic speech recognition result b.

Figures 2, 3 and 4 are "A" and "A" respectively.
These are the calculation results of the Euclidean distance between the first half of the speech interval and the average feature pattern for monosyllabic sounds such as "sa" and "ta", and "FRAME"
indicates the frame number for each syllable, “DIST”
represents the Euclidean distance value. Furthermore, in these figures, the threshold value is set to 0.2 and the frames where the count value is 5 are underlined. From the beginning of the word to the underlined frame, that is, "A" (Figure 2)
In the case of , 5 frames from the beginning of the word, "sa" (Figure 3)
In the case of , 20 frames from the beginning of the word, "ta" (Figure 4)
In the case of , the feature pattern is 15 frames from the beginning of the word, so compared to the case where the feature pattern is for all frames from the beginning to the end of the word, the amount of feature patterns is smaller, and therefore the storage capacity required for the device is smaller. This not only reduces the amount of processing required to identify input speech, but also allows recognition results to be obtained in a shorter time.

Figures 5, 6, and 7 are "A" and "A", respectively.
For monosyllabic sounds such as “sa” and “ta”, the threshold value is
It shows the time change of the spectral envelope based on linear prediction when the count value is 0.2 and 5, and the spectral characteristics of the consonant part and the spectral characteristics of the vowel part of each syllable can be easily grasped.

As described above, according to this embodiment, the monosyllabic speech a
The feature pattern generator 4 calculates a feature pattern from the beginning of the word a, and the average feature pattern generator 3 calculates an average feature pattern from the end of the monosyllabic speech a, and calculates the distance between the feature pattern and the average feature pattern for the word. is calculated by the inter-feature pattern distance calculation unit 6, and when the distance is smaller than a predetermined threshold of 0.2 five consecutive frames by the threshold value determination unit 7, the analysis operation for determining the feature pattern is terminated, and the feature pattern is determined by the threshold determination unit 7. The voice recognition unit 9 recognizes the voice from the characteristic pattern in the pattern holding unit 5, thereby making it possible to perform voice recognition in a short time.

In addition, in this example, the characteristic pattern is
We used the Euclidean distance using the PARCOR coefficient and the distance between feature patterns as a measure, but the feature pattern may be any pattern that can express the characteristics of the input voice, such as the output of a filter bank, and the distance measure may also be used. For example, city distance or cosh
It is also effective when using various distance measures such as scale.

Furthermore, if the distance between the feature patterns is not smaller than the threshold in the threshold determination section 7, a count clear signal for clearing the count of the counting section 8 is output from the threshold determination section 7 to the counting section 8 to set the threshold value. If the distance between feature patterns is smaller than the threshold value in the determination unit 7, a count value increase signal is sent to increase the count value of the counting unit 8 by 1.
If the threshold determination unit 7 outputs the data to the counting unit 8, the feature patterns up to the point in time when frames with the inter-pattern distance smaller than the threshold are consecutive for the number of frames corresponding to the count value are used to calculate the Syllable speech input may also be identified.

As described above, the present invention obtains all the feature patterns of the non-stationary part in a part such as a consonant part that has a temporally non-stationary feature pattern, and also obtains all the characteristic patterns of the non-stationary part, such as a vowel part, which is more constant and has a longer duration than the consonant part. By configuring so that not all feature patterns are obtained for long parts of The amount of patterns can also be reduced, the high-speed processing required of the device can be eased, and at the same time the storage capacity can be reduced, which has great industrial value.

[Brief explanation of drawings]

FIG. 1 is a block diagram of a monosyllabic speech recognition device according to an embodiment of the present invention, and FIGS.
The figure shows the calculation results of the Euclidean distance between the feature pattern and the average feature pattern, and FIGS. 5, 6, and 7 are waveform diagrams of the spectrum envelope obtained by linear prediction. 1... Voice holding section, 2... Voice section detection section, 3
...Average feature pattern generation section, 4... Feature pattern generation section, 5... Feature pattern holding section, 6...
Inter-feature pattern distance calculation unit, 7...threshold determination unit,
8...Counting section, 9...Speech identification section.

Claims (1)

[Scope of Claims] 1. A speech section detecting means for detecting a monosyllabic speech section, a speech holding means for holding the detected monosyllabic speech, and a speech section detecting means for detecting a monosyllabic speech section; and an average feature pattern that generates an average feature pattern by averaging feature patterns of a plurality of specific frames from the end of the monosyllabic speech held by the sound holding means. generating means; inter-feature pattern distance calculating means for sequentially calculating the distance between the feature patterns of the average feature pattern and the voice feature pattern; and comparing the distance between the feature patterns with a predetermined threshold to determine the magnitude thereof. a threshold value determining means for determining the distance between the feature patterns; a counting means for counting the number of frames in which the distance between the characteristic patterns is smaller than the threshold value, and generating a signal when the counted value reaches a predetermined value; feature pattern holding means for holding the feature pattern generated by the feature pattern generation means at regular intervals until the signal generated by the signal arrives; and a voice for identifying input monosyllabic speech using the feature pattern of the feature pattern holding means. identification means, the inter-feature pattern distance calculation means sequentially calculates a distance from the beginning of the monosyllabic speech interval detected by the speech interval detection means, and the threshold determination means calculates the distance between the distance and the threshold. The counting means counts the number of frames determined to be small by the threshold determining means, and when the number of frames counted by the counting means reaches the preset threshold, the signal is A monosyllabic speech recognition device configured to recognize an input monosyllabic speech using a characteristic pattern up to that point when a monosyllabic speech occurs.