JPS6147439B2 - - Google Patents

Description

[Detailed description of the invention] The present invention relates to a speech recognition device.

In recent years, analytical research on human speech has progressed, and many methods are being developed for extracting parameters representing the characteristics of speech from speech. A speech recognition device using such speech parameters has been realized.

It is desired that this audio parameter has as small an amount of information as possible and is effective for recognition.Currently, 10 to 20
A detection output value from a channel-like bandpass filter and a 10th-order partial autocorrelation coefficient, which is a type of linear prediction coefficient, are used.

Such a speech recognition device selects the most similar reference speech parameters by comparing the speech parameters of the input speech with the speech parameters of a large number of standard reference speeches stored in advance, and selects the most similar reference speech parameters. The voice is recognized as the input voice at this time. However, in human speech, even if the same word is uttered repeatedly by the same speaker, there is considerable variation in the duration of the utterance of the word, so in order to obtain a practical recognition rate, it is necessary to calculate the comparison of speech parameters. It is necessary to perform time axis alignment of the audio parameters of both the input audio and the reference audio.

Currently, such time axis matching means performs speech recognition on a word-by-word basis, and includes linear matching means that linearly expands and contracts the time axis according to the time length of the input speech to match the time length of the reference speech, and There is a nonlinear matching means that performs nonlinear expansion and contraction to minimize the squared error between both patterns using a systematic programming method.
When this non-linear matching means is used in a speech recognition device, the recognition rate can be improved more than when using the above-mentioned linear matching means, but since the amount of calculation processing is enormous, real-time processing cannot be performed with a normal microprocessor. Have difficulty.

In view of the current situation, the present invention has been made by focusing on the fact that the time length fluctuation of human speech is almost always dependent on the time length fluctuation of the voiced sound section. The present invention provides a speech recognition device that achieves time axis alignment between input speech parameters and reference speech parameters in each of the speech and unvoiced sections.

FIG. 1 is a block diagram showing the configuration of the speech recognition device of the present invention, in which 1 is a microphone that converts input speech into a speech signal, 2 is an amplifier that amplifies the speech signal, 3 is a voice feature extraction means for extracting characteristic parameters of input voice from the voice signal obtained by the amplifier 2; A plurality of band-pass filters 4, 4... provided, a multiplexer 5 that sequentially selects and extracts each frequency band component from the band-pass filters 4, 4..., and digitally converts each frequency band component value from the multiplexer 5. and an A/D converter 6 for converting input audio parameters into quantitative input audio parameters. Reference numeral 7 denotes a determining means for determining voiced and unvoiced sections of input speech based on the audio signal from the amplifier 2 to obtain a voiced and unvoiced pattern; The voice feature extraction means 3 together with the voiced and unvoiced patterns
This is to temporarily store the input audio parameters obtained from the . Reference numeral 9 denotes a reference voice storage means, which stores a plurality of pairs of voiced and unvoiced patterns of the reference voice and reference voice parameters in association with each other. Reference numeral 10 denotes a comparison means for comparing the voiced and unvoiced patterns of the input voice storage means 8 and the voiced and unvoiced patterns of the reference voice storage means 9 on the time axis, and compares the time of the voiced sound section or the unvoiced section of the input sound. The ratio of the time length of a voiced section or a voiceless section of the reference speech to the length of the reference speech is derived for each corresponding section. Reference numeral 11 denotes a time axis matching means which, based on the comparison result from the comparing means 10, expands or expands proportionally to the time axis in each of the voiced sound section and the unvoiced sound section of the input audio parameters of the input audio storage means 8. This is a shortening. Reference numeral 12 denotes a recognition processing means, which performs calculations on the reference voice parameters of the reference voice parameters and the input voice parameters time-axis aligned according to the reference voice parameters by the time-axis alignment means 11 with multiple references. This is executed on the parameters, and the reference voice indicated by the reference parameter determined to be most similar is recognized as the current input voice, and a signal corresponding to the input voice is output.

FIG. 2 shows an example of means for determining voiced/unvoiced speech used in such a speech recognition apparatus of the present invention. In the figure, 13 is a high-bass filter that extracts only the high-frequency components of the audio signal, 14 is a comparator that compares the high-frequency component value from this filter 13 with a predetermined threshold voltage, and 15 is a high-pass filter that extracts only the high-frequency components of the audio signal. A counter 16 counts the number of zero-crossing waves obtained, and a discriminator 16 discriminates the magnitude of the count value obtained by the counter 15 per unit time.The determining means having such a configuration can be used as a voiced/unvoiced determining means based on zero-crossing waves. It is common.

A case in which numbers are recognized using such a speech recognition device of the present invention will be specifically described. In this case, the recognition operation becomes possible by storing standard voice parameters from 0 to 9 and voiced/unvoiced patterns in the reference voice storage means in advance. During this recognition operation, when the speaker utters, for example, ichi into the microphone 1, this input voice signal is amplified by the amplifier 2 and input to the voice feature extraction means 3, and the voiced/unvoiced determination means 7. will be introduced in The voice feature extraction means 3
In this case, each frequency component of the input audio, i.e., obtained from a plurality of bandpass filters 4 each having a different passband, is sequentially outputted via a multiplexer 5 and an A/D converter 6, and is converted into digital input audio parameters. is obtained. On the other hand, in the above-mentioned determination means 7, as shown schematically in FIG. For example, an H level binary signal is output for a certain (ch) section. ichi expressed in the order of (presence) (absence) (presence) obtained from the determination means 7 in this way
Both the voiced and unvoiced pattern of the input voice and the input voice parameter ichi obtained from the voice feature extraction means 3 are stored in the input voice storage means 8. After that, the voiced and unvoiced patterns are first transmitted from the input voice storage means 8 to the comparison means 10,
In the comparison means 10, the numbers 0 to 9 in the reference voice storage means 9 are compared with the voiced and unvoiced patterns of the reference voice. At this time, first, the order and number of the voiced and unvoiced sections of the voiced and unvoiced pattern are compared. 1 (ICHI) and 6 (ROKU) with . Next, among these selected reference voices, regarding reference voice 1 (ICHI) having voiced/unvoiced patterns and reference voice parameters as schematically shown in FIGS. 3d and 3e, the first Voiced interval (i)
The time length ratio of the first voiced sound section (I) of the reference sound is derived, and the same time length ratio is also calculated for the unvoiced sound section (ch) (CH) and the second voice sound section (i), (I). Derive. Based on the results of these comparisons, the time axis matching means 11 adjusts the input voice parameter 1 by linearly expanding or contracting the time axis for each of the voiced sound section and the unvoiced sound section, as schematically shown in FIG. 3c. and set ichi's input audio parameters and
Time axis alignment with reference audio parameters of ICHI is performed. The input audio parameters of ichi and the reference parameters of ICHI, which are time-aligned in this manner, are subjected to arithmetic processing in the recognition processing section 12, and the results are temporarily stored. The other 6 selected by the comparison means 10 processes such audio parameters.
The same process is performed for the reference parameter of (ROKU), and the input voice ichi is recognized as 1 based on the arithmetic processing result in this recognition processing means 12, and a signal corresponding to this 1 is output from the recognition processing means 12. Ru.

As is clear from the above description, the speech recognition device of the present invention includes a comparison means for comparing voiced and unvoiced patterns extracted from input speech and voiced and unvoiced patterns of reference speech, and a comparison means for comparing voiced and unvoiced patterns extracted from input speech and voiced and unvoiced patterns extracted from input speech based on the comparison result. A time axis matching means for aligning the input audio parameters and the reference audio parameters in the corresponding voiced sound sections and unvoiced sound sections, respectively; Since the present invention includes recognition processing means for recognizing input speech by executing the calculations for a large number of reference parameters, effective time axis alignment of speech parameters is possible. Therefore, a higher recognition rate can be obtained than when using the conventional linear time axis alignment means, and the processing time in the time axis alignment means can be significantly shortened than when using the conventional nonlinear time axis alignment means. I can do things.

Further, the speech recognition device of the present invention has two comparison processing means, the above-mentioned comparison processing means 10 for voiced and unvoiced patterns, and the above-mentioned recognition processing means 12 for voice parameters. In the comparison means 10, which is the first-stage comparison processing means, it is extremely difficult to exclude from the recognition target a reference speech having a voiced/unvoiced pattern that is significantly different in form from the voiced/unvoiced pattern of the input speech at the time of this comparison. It is valid. In other words, since a voiced and unvoiced pattern is composed of a combination of voiced and unvoiced sections with different time lengths, the order and number of voiced and unvoiced sections of the input speech are determined by the allowable order and number of voiced and unvoiced sections when comparing both voiced and unvoiced patterns. As a standard, a rough comparison process is performed that ignores the time length of these sections, a reference pattern to be recognized is selected, and the recognition processing means 12, which is the second stage comparison processing means, performs voice recognition. Parameters are compared and calculated for final recognition. Therefore, the amount of calculation processing in the recognition processing means 12 is reduced, leading to a reduction in processing time.

As a result of the above effects, the present invention realizes a speech recognition device that has a high recognition rate and a short response time, and is highly useful.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the configuration of the speech recognition device of the present invention, FIG. 2 is a block diagram showing the configuration of the determination means used in the device of the present invention, and FIG. 3 is a schematic diagram showing voice parameters and voiced and unvoiced patterns. 3 is a voice feature extraction means, 7 is a determination means, 8 is an input voice storage means, 9 is a reference voice storage means, and 10
Reference numeral 11 indicates a comparison means, 11 a time axis matching means, and 12 a recognition processing means.

Claims (1)

[Scope of Claims] 1. A microphone that converts input speech into an audio signal, an audio feature extraction means that extracts feature parameters of the input audio from the audio signal, and a voiced sound section of the input audio based on the audio signal. a human voice storage means for temporarily storing the voiced and unvoiced patterns obtained from the judgment means together with the input speech parameters obtained from the speech feature extraction means; a reference voice storage means storing a plurality of voiced and unvoiced patterns in association with each other; and a time period of each voiced or unvoiced section corresponding to the voiced and unvoiced pattern of the input voice storage means and the plurality of voiced and unvoiced patterns of the reference voice storage means. A comparison means that performs a comparison with respect to the axis, and a linear time axis matching between the input voice parameter of the input voice storage means and the reference voice parameter of the reference voice storage means based on the time length comparison result from the comparison means. a time axis matching means that performs the operation for each voiced or unvoiced section, and calculations between input audio parameters and reference audio parameters that are time axis aligned by the time axis matching means are performed for a large number of reference audio parameters; and recognition processing means for recognizing a voice of a reference voice parameter determined to be most similar to the input voice parameter as the input voice. 2. In the speech recognition device according to claim 1, the comparison means is provided with a predetermined acceptance criterion, and the voiced and unvoiced pattern of the reference speech storage means and the voiced and unvoiced pattern of the input speech storage means are A speech recognition device characterized in that, if the difference in voiced and unvoiced patterns deviates from the acceptance criteria, reference speech parameters associated with this voiced and unvoiced pattern are excluded from calculation targets for input speech parameters.