JPS6225798A - 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] 'E, -D Win the 11th of November! φ日■ [Industrial Application Field] The present invention relates to a speech recognition device that is applied to, for example, recognizing the speech of a specific speaker word by word.

[Summary of the invention]

The present invention calculates a tendency value for correcting the spectral tendency that fluctuates due to various causes in a speech recognition device, and flattens the spectral tendency based on this tendency value, thereby eliminating individual differences between speakers. This is intended to improve the recognition rate without being affected by surrounding noise, etc.

[Conventional technology]

As a conventional audio Elm& device, one is known that is composed of, for example, a microphone as an audio input section, a preprocessing circuit, an acoustic analyzer, a feature data extractor, a registered pattern memory, a pattern matching determiner, 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 audio signal. feed the analyzer.

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 N-dimensional hector trajectory length 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 equally divided by the predetermined number of divisions necessary to extract the 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 registered feature data processor (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 human 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.

[Problem that the invention seeks to solve]

However, the tendency of the frequency spectrum of a voice signal changes greatly due to individual differences between speakers and the inclusion of surrounding noise, and unless this tendency is normalized, the recognition rate will be extremely low.

For example, the frame data shown in FIG. 5A is deformed by noise having a spectral tendency as shown in FIG.
Assume that the frame data is as shown in Figure C. When the pattern matching determiner determines the distance between the frame shown in FIG. 5A and the frame shown in FIG. 5C,
The inter-frame distance is calculated as a large value, and the matching distance is assumed to include a large error, increasing the possibility that it will be misrecognized. For this reason, it has been proposed to flatten (normalize) the spectral trend by correcting the spectral trend fluctuation so that it is not influenced by individual differences between speakers or ambient noise.

For example, methods have been proposed in which the spectral tendency is estimated and normalized using a linear function using the least squares method, and a method in which normalization is performed using a correction function that is partially averaged between a predetermined channel width. Not only is the calculation complicated, it cannot be applied when the trend is curved, and in the latter case,
This method has the disadvantage that it cannot be applied when the spectral envelope is smooth.

Accordingly, an object of the present invention is to provide a speech recognition device having means capable of accurately normalizing any spectral tendency simply and quickly.

[Means for solving problems]

The present invention provides a speech recognition device in which time-series data of N-channel frequency spectra is converted into N-channel frequency spectra and is input, in which all channels lower than a predetermined channel are A first average value of the spectral data is calculated, and a second average value of the spectral data up to all channels higher than a predetermined channel is calculated, and the average value of the first average value and the second average value is calculated. The value is calculated as a trend value corresponding to a predetermined channel, and the trend value is offset from the spectral data to obtain trend 11 layered smoothness data to recognize the input audio signal. It is a voice J, knowledge device that is characterized by the following.

[Operation] A spectral tendency normalizer 6 is provided as a means for normalizing the spectral tendency. 1≦n≦N), and the average value of spectral data from a predetermined channel n to the largest channel N is determined: J +1. , the average value of each average value is further calculated and used as a trend value for a predetermined channel n, and a subtraction process is performed between the spectrum data of each channel and the corresponding trend value to calculate Q.
spectral trends are flattened.

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

Part 1) shows a practical example of this invention.
In the figure, l indicates a microphone as an audio input section.

An analog audio signal from microphone 1 is supplied to filter 2 . Filter 2 is, for example, a force filter.

It is a low-pass filter with a turn-off frequency of 7.5KHz,
The audio signal is filtered in a filter 2 to limit the band below 7.5 Hz, which is required for speech recognition, and this audio signal is supplied to an A/D converter 4 via an amplifier 3.

For example, the A/D converter 4 has a sampling frequency of 12.
5K)lz 8-bit A/D converter, the audio signal is analog-to-digital converted in the A/D converter 4 to become an 8-bit digital signal, and the audio signal is converted into an 8-bit digital signal.
supplied to

The acoustic analyzer 5 converts the audio signal into a frequency spectrum (Q) to generate, for example, an N-channel spectral data string. In the acoustic analyzer 5, the audio signal is converted into a frequency spectrum through arithmetic processing, and, for example, a spectral data string whose representative value is 131 wave numbers of N fil+1, which are spaced 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. Then, every unit time (frame cycle! IJl), a thousand spectral data strings are output as one frame data. That is, for each frame, the audio signal is cut out as parameters expressed by N-dimensional vectors, and is supplied to the spectral trend converter 6.

For example: , the frame corresponding to the end of the voice section is 1
In this case, as shown in Figure 2, each channel 1
Frame data constituted by data from 1 frame to 1 frame are supplied to the spectral tendency normalizer 6.

In the spectral tendency normalizer 6, the tendency [F normalization process] of the spectral data is performed for each sequentially supplied frame data. That is, regarding the spectral data of each channel constituting the frame data (the tendency value F7 for correcting intermediate fluctuations is calculated by the following 2).

In other words, from channel 1 to predetermined channel n (1≦n
≦N), the average value of the spectrum data from the predetermined channels I and N to the maximum channel N is determined, and the average value is determined for the history of each average value, and around 1 It is considered to be straight forward F1. By subtracting the spectral data of each channel and the corresponding trend value F1, the spectral trend is flattened, and the spectral trend is unaffected by individual differences between speakers and surrounding noise. Normalized. l
Trend normalization processing is similarly performed on all frames from frame to frame 1, and the trend-normalized frame data is supplied to the feature data extractor 7.

In the feature data extractor 7, the distance between the matching frame data is calculated. For example, the absolute value of the difference in spectrum data for each channel is determined, and the sum of the values is determined as the interframe distance.

In the history, the sum of the interframe distances is determined, and the trajectory length in N-dimensional hectares from the start frame to the end frame of the audio signal is determined. Then, in the case of the longest speech with the largest number of words, the trajectory length is divided equally by a predetermined number of divisions necessary to extract features, and only the frame data corresponding to the division points are extracted as feature data, and the speaker's The time axis is normalized and output so as not to be affected by variations in the rate of sound generation.

At the time of registration, this feature data is supplied to the registered pattern memory 8 and stored as a registered feature data block. At the time of recognition, the input audio signal undergoes the above-mentioned processing, is converted into an input feature data processor, and is supplied to the pattern matching determiner 9, where patterns are determined between the input feature data block and all registered feature data blocks. Matching is performed.

In the pattern matching determiner 9, the interframe distance is determined between the frame forming the input feature data block and the frame forming the registered feature data block to be compared, and the sum of the distances is determined as the matching distance. It is said that Then, a word corresponding to the registered feature data block that is determined to be the smallest and sufficiently close among the matching distances determined for all registered feature data blocks is output as a recognition result.

The operation of the spectral tendency normalizer 6 in the embodiment of the present invention described above will be explained with reference to the flowchart shown in FIG.

Frame data is sequentially supplied from the acoustic analyzer 5 to the spectral tendency normalizer 6, and steps ■ to ■ are performed for each frame.
processing is performed.

First, a variable n indicating a channel number is initialized to 1 (step ■). In step ■, channel 1
A correction function is calculated, and a trend value F, which corrects the trend fluctuation, is obtained from 2(N+1-1)·l.

Then, normalization processing is performed in step (2), and the trend value F1 is subtracted from the channel 1 spectrum data S1, and the result of this subtraction is taken as the channel I spectrum data S.

In step ■, a variable n(-
1) is compared with the maximum number of channels N, and n is incremented (n=2) (step ■), and the calculation process regarding channel 2 is started.

The trend value F2 for correcting the trend fluctuation is determined by 2 1 + 1-2) 2 (step ■), the trend value F2 is subtracted from the spectrum data S2 of channel 2, and the result of this subtraction is taken as the spectrum data S2 of Nayannel 2. Ru. (
Susie Seop ■).

Further, while n is incremented, the processes of steps ■ to ■ described above are repeated, and the trend value F7 for the predetermined channel is determined by 2 (N+1-n) ・n (step ■), and the spectrum data S of the predetermined channel is obtained. , , (intermediate (direct F, , is subtracted from
This M calculation result is the spectrum data S1 of the predetermined channel.
It is said that When the variable n indicating the channel number is set to the maximum number of channels N, and the trend normalization process is performed on the maximum channel N, the calculation regarding one frame is completed.

For example, it is composed of 16 spectral data of channels 1 to 16 as shown in FIG.
16, 17, 12, 14, 18, 16, 12°10,
6.12. 9. 8. 6. 5), the trend values F, from channel 1 to channel 16 draw the curve shown in Figure 4B, and the trend normalized spectrum data is based on the trend value F1. It is flattened and is shown in the 41st piece C. In this way, the trend normalization process is performed on all frames, and the spectrum around 1 is flattened while preserving unique characteristic spectral data.

In one embodiment of the present invention, the spectral tendency normalizer 6 has been described as being provided before the feature data extractor 7, but the spectral tendency normalizer 6 may be provided after the feature data extractor 7. It's good as well. Further, the present invention is not limited to a hard-wired configuration, and processing may be performed by software using a microcomputer or a microprogram method.

〔Effect of the invention〕

In this invention, a spectral tendency normalizer is provided as a means for normalizing the spectral tendency, and in the spectral tendency normalizer, channels l to predetermined channels n (1≦n≦N )
The average value of the spectrum data up to
The average value of the spectrum data from a predetermined channel to the maximum channel N is calculated, and the average value of each average value is further calculated to obtain a trend value for the predetermined channel number N, and the trend value corresponding to the spectrum data of each channel is calculated. The spectral tendency is flattened by performing subtraction processing between the two.

Therefore, according to the present invention, it is possible to accurately normalize any spectral tendency with a thin structure and at high speed, and the calculation processing time is shortened and the E'2-Rh rate is improved.

Further, in conventional voice recognition 3all devices, there are restrictions on identifying the speaker and preventing the mixing of foreign sounds such as ambient noise, but this invention eliminates the need for this.

[Brief explanation of drawings]

Fig. 1 is a block diagram of the overall configuration of an embodiment of the present invention, Fig. 2 is a route map used to explain the data structure of time-series frame data in an embodiment of the invention, and Fig. 3 is a route diagram of this embodiment. Flowcharts of FIGS. 4A, 4B, and 4C used to explain the operation of the spectral tendency normalizer in an embodiment of the invention are used to explain the operation of the spectral tendency normalizer in an embodiment of the invention. Route Map FIG. 5 is a route map used to explain the conventional technology. Explanation of main symbols in the drawings 1 - Microphone, 5: Acoustic analyzer. 6: Spectral trend normalizer. 7: Feature data extractor. 8: Registered pattern memory. 9; Pattern matching judge. Agent Patent Attorney Tadashi Sugiura Chizen 1 0 oval Figure 1 ■ ~ Itori) 70-10-Yato Figure 3 Figure 4A Figure 4B Figure 4C

Claims (1)

[Scope of Claims] A speech recognition device in which an input audio signal is converted into N-channel frequency spectra, and to which time-series data of the N-channel frequency spectra is input, wherein a predetermined method is provided for the spectrum data of each frame of the time-series data. A first average value of the spectral data of all channels lower than the channel is calculated, and a second average value of the spectral data of all channels higher than the predetermined channel is calculated, and the first average value of the spectral data of all channels higher than the predetermined channel is calculated. and the second average value is calculated as a trend value in the predetermined channel, and the trend value is offset from the spectrum data to obtain trend-normalized spectrum data to obtain the input audio signal. A speech recognition device characterized in that it recognizes speech.