JPS63121100A - Feature pattern extraction for 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 [Industrial Field of Application] The present invention relates to a feature pattern extraction method in a speech recognition device that is applied, for example, to recognizing speech of unspecified speakers word by word.

[Summary of the invention]

The present invention provides a feature pattern extraction method in a feature extraction unit of a speech recognition device.
” area and extract the number of channels existing in each area as a feature value to easily normalize translational fluctuations in the frequency axis direction. This is intended to improve the recognition rate without being influenced by differences.

[Conventional technology]

Japanese Patent Application No. 59-1 previously proposed by the applicant
The speech recognition device shown in the specification of No. 06177 includes a microphone as a speech input section and a preprocessing circuit.

It is composed of an acoustic analyzer, a feature data extractor, a registered pattern memory, a pattern matching judger, etc.

This speech recognition device limits the audio signal input from the microphone to the band required for speech recognition in a preprocessing circuit, converts it into a digital audio signal using an A/D converter, and converts this digital audio signal into an acoustic analyzer. supply to.

Then, in the acoustic analyzer, the audio signal is converted into a frequency spectrum, and the frequency spectrum is normalized using N frequencies as representative values, for example, at regular intervals on the logarithmic axis. Frame data constituted by the data is supplied to a feature data extractor.

The feature data extractor calculates the distance between adjacent frame data, calculates the trajectory length of the N-dimensional vector from the start frame to the end frame of the audio signal by summing the distances between each frame, and extracts the longest audio with the largest number of words. In this case, the trajectory length is divided into sections using a predetermined number of divisions necessary to extract features, and only the frame data corresponding to the division points are extracted as feature data, thereby adjusting for fluctuations in the rate of speech generation of the speaker. Normalize and output the time axis so that it is not affected.

When registering this feature data, it is supplied to the registered pattern memory and used as a registered feature data block (standard pattern).
At the time of recognition, the input audio signal is processed as described above to become an input feature data block (input cover pattern) and supplied to a pattern matching determiner. A pattern matching determiner performs pattern matching between the input feature data block and the registered feature data block.

The pattern matching determiner calculates the inter-frame distance between the frame data constituting the registered feature data block and the frame data constituting the input feature data block, and uses the sum of the inter-frame distances as the matching distance, The matching distance is similarly calculated for the feature data block, and the word corresponding to the registered feature data block for which the matching distance is the minimum and the distance is determined to be sufficiently close is output as a recognition result.

However, in conventional speech recognition devices, the frame data output from the acoustic analyzer is directly stored in the registered pattern memory as a registered feature data block via the feature data extractor, so the amount of memory in the registered pattern memory is enormous. There were some serious problems. At the same time, there was also the problem that during pattern matching, the calculation processing time was long depending on the amount of data. ゛For this reason, each of the spectral data that makes up the frame data is binarized, and the registered pattern memory is The applicant of the present application has proposed a speech recognition device (Japanese Patent Application No. 166191/1982) which shortens the matching processing time by reducing the capacity.

This speech recognition device calculates a tendency value to correct the tendency of the spectrum that fluctuates due to various causes, and flattens the tendency of the spectrum based on this tendency value to compensate for the influence of individual differences between speakers and surrounding noise. After the frame data is normalized to avoid any distortion, each piece of spectral data constituting the frame data is binarized, and pattern matching is performed based on the obtained binary pattern.

[Problem that the invention seeks to solve]

However, it is known that the time-series frequency spectrum of an audio signal generally fluctuates not only in the time axis direction but also largely in the frequency axis direction due to the speaking style of speakers, differences among speakers, and the like. For this reason, in speech recognition devices that use conventional feature pattern extraction methods that involve only time-axis normalization processing, it is necessary to specify the speaker's speaking style evenly, or use a multi-template template that prepares multiple standard patterns. By using this method, fluctuations in the frequency axis direction were dealt with. However, this speaking style regulation when inputting voice is not practical, and when the multi-template method is used, the amount of registered pattern memory becomes enormous.
Along with this, there is a drawback that the processing time becomes longer. - By the way, variations in the frequency axis direction due to differences between speakers, such as men, women, children, and adults, generally involve parallel shifts on a logarithmic scale, and if such variations can be normalized, it can be It is possible to improve the recognition rate of speech targeted at a specific speaker. Therefore, there is a need for a feature pattern extraction method that can easily normalize translational fluctuations in the frequency axis direction.

Therefore, an object of the present invention is to provide a feature pattern extraction method in a speech recognition device that can easily normalize translational fluctuations in the frequency axis direction and improve the recognition rate for unspecified speakers. There is a particular thing.

[Means for solving problems]

The present invention includes a step (binarizer 12) of comparing spectral patterns of input audio signals with a predetermined threshold value to obtain a binary spectral pattern; Alternatively, a step (area distance feature extractor 13) of calculating the consecutive number of "0" to obtain a pattern matrix, and normalizing the values included in a predetermined column of the pattern matrix corresponding to the ends of the binary spectral pattern. (frequency axis normalizer 15).

[Effect]

Matrix XM of M rows and N columns output from the binarizer 12
A binary spectral pattern expressed by N is supplied to the region distance feature extractor 13, and a series of "1"s in the channel direction are
and the area of “0”, and each area (1≦i51.
i is an integer) is extracted as a feature value. Through this extraction process, an initial feature pattern expressed by a matrix XNI of M rows and I columns is formed, and this initial feature pattern is supplied to the time axis normalizer 14.

In the time axis normalizer 14, normalization processing is performed along the time series trajectory, and the initial feature pattern is made into a feature pattern that is not affected by fluctuations in the time axis direction expressed by a matrix XJI with J rows and ■ columns, and this feature pattern is supplied to the frequency axis normalizer 15.The frequency axis normalizer 15 determines the minimum feature value among the feature values in the (i-1) column of the feature pattern, and The feature value that is determined to be the minimum is subtracted from each feature value, and each obtained value is used as a new feature value for the column (i-1) to form a feature pattern that is not affected by fluctuations in the frequency axis direction. , pattern matching is performed based on this feature pattern.

〔Example〕

a. Description of the overall structure and operation of the speech recognition device An embodiment of the present invention will be described below with reference to the drawings. Figure 1 shows
This figure shows an outline of an example of a speech recognition device used to put the present invention into practice, and the present invention relates to the processing of the feature extraction unit shown by 6 in FIG.

In FIG. 1, l indicates a microphone as an audio input section. An analog audio signal from microphone 1 is supplied to filter 2 . The filter 2 is, for example, a low-pass filter with a cutoff frequency of 7.5 klLz, and the audio signal is limited in the filter 2 to a band of 7.5 kHz or less required for speech recognition.
This audio signal is supplied to an A/D converter 4 via an amplifier 3.

The A/D converter 4 has a sampling frequency of 16.5, for example.
This is a kHz 8-bit A/D converter, and the audio signal is
The D converter 4 performs analog-to-digital conversion into an 8-bit digital signal, which is then supplied to the acoustic analyzer 5.

The acoustic analyzer 5 converts the audio signal into a frequency spectrum and generates, for example, an N-channel spectrum data string. In the acoustic analyzer 5, the audio signal is converted into a frequency spectrum through arithmetic processing, and a spectral data string whose representative values are, for example, N frequencies at regular intervals on the logarithmic axis is obtained. Therefore, an audio signal is expressed by the magnitude of discrete frequency spectra of N channels, and a spectral data string of N channels is outputted as one frame data every unit time (frame period). That is, the audio signal is cut out as a parameter expressed by an N-dimensional vector every frame period, and is supplied to the feature extraction section 6.

For example, if the frame corresponding to the end of the audio section is M, as shown in Figure 2, the frame data each consisting of spectrum data of channel 1 to channel N is chronologically arranged from frame 1 to frame M. It is supplied to the feature extraction section 6.

The feature extraction unit 6 performs various processes such as spectral tendency normalization, binarization, region distance feature extraction 1 time axis normalization, and frequency axis normalization to form a feature pattern, as will be described in detail later. The feature pattern formed in the feature extraction section 6 is supplied to the mode switching circuit 7.

When this characteristic pattern is registered, it is supplied to the registered pattern memory 8 via the mode switching circuit 7 and stored as a standard pattern. Further, during recognition, the input audio signal is converted into a feature pattern by the above-described processing, and this feature pattern is supplied to the pattern matching determiner 9 as an input pattern. Then, the standard pattern in the registered pattern memory 8 to be compared is supplied to the pattern matching determiner 9, and pattern matching is performed between the input cover pattern and all the standard patterns to be compared. .

In the pattern matching determiner 9, inter-frame distances are determined between the frames constituting the input cover pattern and the frames constituting the standard pattern to be compared, and the sum of these distances is taken as the matching distance. Then, a word corresponding to a registered pattern that is determined to be the smallest and sufficiently close among the matching distances found for all registered patterns to be compared is output as a recognition result.

b. Description of feature pattern extraction method in feature extraction unit FIG. 3A shows an example of the configuration of the feature extraction unit 6 in the speech recognition device described above. 6 is a spectral tendency normalizer 11. Binarizer 12. It is composed of a region distance feature extractor 139, a time axis normalizer 14, and a frequency axis normalizer 15.

Time-series frame data from the acoustic analyzer 5 is supplied to the spectral trend normalizer 11, and the spectral trend normalizer 11 performs trend normalization processing on the spectral data for each sequentially supplied frame data. ,
The trend value for correcting trend fluctuations regarding the spectrum data constituting each frame data is the average value of the spectrum data from channel 1 to a predetermined channel n (1≦n≦N, n is an integer), and the maximum value from the predetermined channel n. It is obtained by multiplying the average value of the spectrum data up to channel N by an appropriate coefficient. Subtraction is performed between the trend value determined for the spectrum data of each channel and the corresponding spectrum data, and the spectrum trend is flattened so that it is not influenced by individual differences between speakers, surrounding noise, etc. Spectral data orientation is normalized. Trend normalization processing is similarly performed on all frames, and the trend-normalized frame data is supplied to the binarizer 12.

In the binarizer 12, 8-bit spectral data constituting frame data is processed with a threshold value set to an appropriate value so that the formant part of the spectral envelope expressed by one frame data is 1. For example, the spectrum data is compared with a threshold value set to an appropriate value, and the spectrum data with a value greater than the threshold value is set to "1", and the threshold value is The spectrum data having the smaller value is set to "0". This binary spectral pattern formed in the binarizer 12 is supplied to the region distance feature extractor 13.

The region distance feature extractor 13 determines a region of "1" and a region of "0" that are continuous in the frequency axis direction of each frame, that is, in the channel direction, and uses the number of channels existing in each region as a feature value. This extraction process forms an initial feature pattern.

In other words, the channel number on the frequency axis is n (1≦n≦
N, where n is an integer), and the frame number as the time axis is m (1≦m≦M, m is an integer), and the binary spectral pattern is expressed as an M row, N column matrix Xl''N. For each frame (each m rows), determine the continuous "1" area and the continuous "0" area in the channel direction (for example, in the n column direction where the channel number increases), and set the starting area as 1. Area number i from 1 to the terminal area I (1≦i≦I, where i is an integer 5.Then, the number of channels existing in each area is extracted as a feature value, and these feature values are arranged in the order of the area number. In this way, an initial feature pattern expressed by a matrix X- of M rows and I columns is formed.

For example, if each frame is composed of 15 channels from channel 1 to channel 15 as shown in FIG. 4A, and there is a binary spectrum pattern composed of 8 frames, then for each frame A continuous “1” area (shaded area in the figure) and a continuous “0” area (shaded area in the figure)
” region is determined and divided into 6 regions except for exceptional frames. Then, the number of channels existing in each region is extracted as a feature value in a corresponding manner as shown by the arrows in the figure. Therefore, the binary spectral pattern shown in FIG. 4A expressed by an 8 rows and 15 columns matrix is taken as the initial feature pattern expressed by an 8 rows by 6 columns matrix as shown in 4-8. The initial feature pattern thus formed in the region distance feature extractor 13 is supplied to the time axis normalizer 14.

The time axis normalizer 14 uses, for example, a feature vector (a series of
I corresponding to the number of regions I of the "1" region and the continuous rOJ region
The initial feature pattern is compressed in the time axis direction by performing normalization processing along the time series trajectory in space (dimensional vector).
or elongate).

For example, in the time axis normalizer 14, the absolute value of the difference between the feature values of corresponding regions of adjacent frames constituting the initial feature pattern is determined, and the sum thereof is taken as the interframe distance between the adjacent frames. Furthermore, the sum of the inter-frame distances is determined, and the trajectory length of the ■-dimensional vector from the starting frame 1 to the ending frame M is determined. Then, the trajectory length is equally divided by a predetermined number of divisions (for example, J) necessary for extracting features. Only frames existing in close proximity to each of the three division points are extracted, and the time axis is normalized and output so as not to be affected by variations in the rate of speech generation of the speaker. Therefore, the initial feature pattern expressed by the matrix XMI with M rows and I columns is
It is compressed (or expanded) into a characteristic pattern expressed by a matrix of columns XJI. The feature pattern formed in the time axis normalizer 14 is supplied to the frequency axis normalizer 15.

The frequency axis normalizer 15 performs normalization processing on translational fluctuations in the frequency axis direction to form a final feature pattern.

The feature values other than the (i=1) column of the feature pattern from the time axis normalizer 14 have already been made into values that are invariant to translational fluctuations in the frequency axis direction through processing in the region distance feature extractor 13. Therefore, the frequency axis normalizer 15 performs calculation processing only on the feature values of the (i-1) column of the feature pattern. For example, for the feature pattern (i=
The minimum value among the feature values in the column 1) is determined, and the feature value determined to be the minimum is subtracted from each feature value in the column (i = 1). Each value obtained by this subtraction process is used as a new feature value for the example (i=1).

For example, if the feature pattern shown in FIG.
4, resulting in a compressed or expanded form), processing is performed only on the leftmost column (i=1) in FIG. 4B, and the one with the minimum feature value is "3".
It is determined that "3" is subtracted from each feature value. Each value obtained by this subtraction process is set as a new feature value for the column (i=1) as shown by the arrow in the figure, and is shown in FIG. 4B.
The characteristic pattern shown in FIG. 4C is taken as the characteristic pattern shown in FIG.

As described above, the characteristic pattern formed in the frequency axis normalizer 15 is supplied to the registered pattern memory 8 at the time of registration, and is supplied to the pattern matching determiner 9 at the time of recognition.

In one embodiment of the present invention, a case has been described in which a region of successive "1" or "0" is determined from channel 1 in the direction of increasing channel number when performing fII range distance feature extraction processing. is a sequence of successive "1" from channel N in the direction of decreasing channel number.
” or “0”, and an initial feature pattern may be formed corresponding to each region.

Further, in the feature extraction unit 6 of one embodiment of the present invention,
Although we have described the case where processing is performed in the order of spectral tendency normalization, binarization, region distance feature extraction 1, time axis normalization, and frequency axis normalization, the configuration of the feature extraction unit 6 is shown in Fig. 3B.
As shown in the figure, a time axis normalizer 14 is installed before the binarizer 12.
For example, a feature vector (an N-dimensional vector corresponding to the number of channels of the acoustic analyzer 5) is subjected to time-axis normalization processing along a time-series trajectory in space, then binarized, and area distance feature extraction processing is performed. The frequency axis normalization process may be performed after the initial feature pattern is formed by performing the above steps. Furthermore, the feature extraction unit 6 is configured as shown in FIG. 3C, and after the binarization process, the time axis normalization process is performed, and then the area distance feature extraction process is performed to form an initial feature pattern, and the frequency axis normalization process is performed. It is also possible to perform conversion processing.

〔Effect of the invention〕

In this invention, a binary spectral pattern expressed by a matrix XMN of M rows and N columns supplied from a binarizer is supplied to a region distance feature extractor, and a region of "1" and "0" continuous in the channel direction is supplied to the region distance feature extractor. is determined and each area (1≦i
≦I, i is an integer) is extracted as the characteristic m value. Through this extraction process, an initial feature pattern expressed by a matrix X with M rows and 1 column is formed,
This initial feature pattern is supplied to a time axis normalizer.

The time axis normalizer performs normalization processing along the time series trajectory, and the initial feature pattern is a matrix XJ with J rows and I columns.
This feature pattern is expressed by I and is not affected by fluctuations in the time axis direction, and this feature pattern is supplied to the frequency axis normalizer. In the frequency axis normalizer, the minimum feature value in the (i = 1) column of the feature pattern is determined, and the minimum feature value is subtracted from each feature value in the (i = 1) column. , and the obtained values are (1=1
) A feature pattern that is not affected by fluctuations in the frequency axis direction is formed as a new feature value for the column, and pattern matching is performed based on this feature pattern.

Therefore, according to the present invention, translational fluctuations in the frequency axis direction can be easily normalized, and fluctuations in the time axis direction can also be easily normalized. This makes it possible to improve the rate.

Further, according to the present invention, there is no need to specify the speaking style during voice input as in the case of conventional speech recognition devices, and there is no need to use a multi-template method.
It is possible to reduce the capacity of the registered pattern memory and to speed up recognition processing, making it possible to provide a small, low-cost, and practical speech recognition device.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of an embodiment of this invention, FIG. 2 is a route diagram showing the structure of time-series frame data output from an acoustic analyzer in an embodiment of this invention, and FIG. 3A is a block diagram of an embodiment of this invention. FIGS. 3B and 3UjJC are block diagrams showing the configuration of the feature extraction unit in one embodiment of the present invention, and FIGS. The figure is a route map used to explain the operation of the feature extraction section in one embodiment of the present invention. Microphone, 5: Acoustic analyzer, 6: Feature extractor, 8: Registered pattern memory, 9: Pattern matching judger, 11 Varnish vector tendency normalizer, 12: Binarizer , 131 region distance feature extractor,
14: Time axis normalizer, 15: Frequency axis normalizer. Agent Patent Attorney Tadashi Sugiura Intellectual & 9.1 Frame Figure 2 Figure 3 A
Figure 3B Figure 3C

Claims (1)

[Scope of Claims] Obtaining a binary spectral pattern by comparing spectral patterns of input audio signals with a predetermined threshold; and a step of comparing spectral patterns of input audio signals with a predetermined threshold; obtaining a pattern matrix by calculating consecutive numbers of , and normalizing the values included in a predetermined column of the pattern matrix corresponding to the ends of the binary spectral pattern. Feature pattern extraction method in speech recognition device.