JPS625299A - 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 is applied to, for example, recognizing the speech of a specific speaker word by word.

[Summary of the invention]

In a pattern matching determiner of a speech recognition device that uses a close-talking microphone or the like as an input section, the present invention provides input vector data that determines the interframe distance of corresponding frames between a registered feature data block and a human feature data block. The matching distance is calculated by multiplying the weighting coefficient obtained by using the weight corresponding to the amount of change in the frequency axis direction of the registered spectrum data and the weight corresponding to the amount of change in the frequency axis direction of the registered spectrum data to strongly bring out the characteristics of the spectral data string. By calculating and performing a matching judgment, the recognition rate is prevented from decreasing even when a large number of similar registered spectrum data sequences are registered in the memory.

[Conventional technology]

The speech recognition device previously proposed by the applicant (
Japanese Patent Application No. 59-106177) is comprised of a microphone as a voice input section, a preprocessing circuit, an acoustic analyzer, a feature data extractor, a registered pattern memory, a pattern matching determiner, and the like.

The audio signal input from the microphone is limited to the band required for speech recognition in the preprocessing circuit, and then the A/D
A converter converts it into a digital audio signal. This digital audio signal is fed to an acoustic analyzer.

In the acoustic analyzer, the audio signal is converted into a frequency spectrum, and the levels of the frequency spectrum are normalized, for example, at regular intervals on a logarithmic axis, to generate discrete frequency spectrum data.

This frequency spectrum data string is output as one frame data every unit time (frame period). That is, one frame of data for each frame period is, for example, N-channel frequency spectrum data, extracted as a parameter expressed by an N-dimensional vector, and supplied to the feature data extractor.

In the feature data extractor, the distance between adjacent frame data is calculated. The absolute values of the differences between the spectral data of the corresponding channels are determined, and the sum thereof is taken as the interframe distance.

Furthermore, the sum of the distances between each frame is determined, and the trajectory length of the N-dimensional vector 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.

When this feature data is registered, it is stored in the registered pattern memory as a registered feature data block (standard pattern). ! ! ! At the time of recognition, the input audio signal undergoes the above-mentioned processing and is made into an input feature data block, which is supplied to a pattern matching determiner, and pattern matching is performed between the input feature data block and the registered feature data block. .

In the pattern matching judger, the matching distance between the input feature data block and the registered feature data block is calculated. For example, one frame data is extracted in the feature data extractor and θ~(1-1) frames A feature data block is constructed. A distance between corresponding frames is calculated between the frame data forming the registered feature data block and the frame data forming the input feature data block.

For example, the first interframe distance is the absolute value as shown in the following equation, where n is the channel number, the spectrum data of the input feature data block is S in, and the spectrum data of the registered feature data block is Rin. Calculated as distance.

Then, the inter-frame distance 5 is calculated for all corresponding frames, and this inter-frame distance is (i−
1-1), that is, the matching distance is determined. Matching distances are similarly determined for other registered feature cheater blocks, and words corresponding to registered feature data blocks for which the matching distance is the minimum and are determined to be sufficiently close are output as recognition results.

[Problem that the invention seeks to solve]

The interframe distance calculation process in the pattern matching determiner of the conventional speech recognition device is performed as the sum of the absolute values of the differences between the spectral data of corresponding frames and channels, as described above.

For example, if there are three frames each composed of 10 spectral data from channels 0 to 9 as shown in FIG. 5A, FIG. 5B, and FIG. 5C, each frame in FIG. The size of the spectral data of the channel is (
10, 14, 12°9, 6. 8. 10.9. 9
．． The size of the spectrum data of each channel of frame A shown in 8) and FIG. 5C is (9, 13, 10, 1
The inter-frame distance DA from frame C indicated by 3, 8, 6° 16.10, 9.9) is DA = 1 + 1 + 2 + 4 + 2 + 2 + 6 + 1 + 0 + 1 = 2
It becomes 0. Also, the size of the spectrum data of each channel in Fig. 5B is (10, 15, 12, 9°6.8, 15
．． The inter-frame distance D3 between frame B- shown in 8.7.7) and frame C shown in FIG.
0, and from the interframe distance, it is determined that frames A and B are equally similar to frame C.

However, in reality, as is clear from FIGS. 5A, 5B, and 5C, frames B and C are more similar than frames A and C, and the interframe distance is Although it is necessary to calculate the frame distance as smaller than the frame open circuit MDA, the conventional interframe distance calculation process does not yield the expected calculation result.
This causes a decrease in the recognition rate.

Therefore, an object of the present invention is to prevent a decrease in recognition rate by multiplying the inter-frame distance by a weighting coefficient to strongly bring out the features of the spectral data sequence and performing matching distance calculation processing during pattern matching. The purpose of this invention is to provide a speech recognition device that can perform the following tasks.

[Means for solving problems]

This invention comprises acoustic analysis means 5 which performs preprocessing necessary for speech recognition such as spectrum conversion on an input speech signal, and feature data extraction means 6 which is supplied with output data of the acoustic analysis means 5 and extracts feature data from the output data. , a memory 7 in which feature data is stored as a standard pattern, an input pattern from the feature data extraction means 6, and a memory 7
The standard pattern read out from the input pattern S! is supplied, and each spectrum data S! A weighting coefficient Wt is determined using a value W3 corresponding to the amount of change in the frequency axis direction of fi and a value WsI corresponding to the amount of change in the frequency axis direction of each spectrum data Rin constituting the standard pattern, and a weight is applied to the interframe distance. This is a speech recognition device characterized by comprising: a pattern matching determination means 8 which performs a distance calculation process between an input pattern and a standard pattern by multiplying it by a coefficient Ws, and determines matching based on the result of the distance calculation process.

[Operation] In the pattern matching determiner 8, the absolute value of the difference between the spectral data of adjacent channels of each spectral data is determined for each corresponding frame between the registered feature data block and the input feature data block, and the difference is calculated. The values are accumulated to determine the weight Ws of the input waste vector data and the weight WII of the registered spectrum data, and the weight W
A weighting coefficient W is calculated using s and a weight W, and the inter-frame distance of the corresponding frame between the registered feature data block and the input feature data block is multiplied by a weighting coefficient W, and matching is performed using the sum of the multiplication results. The distance is calculated,
Matching determination is performed based on this matching distance.

〔Example〕

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

FIG. 1 shows an embodiment of the present invention. In FIG. 1, 1 indicates a microphone as an audio input section. For example, microphone 1 is a close-talk type microphone with less ambient noise. It will be done.

An analog audio signal from microphone 1 is supplied to filter 2 . The filter 2 is, for example, a low-pass filter with a cut-off frequency of 7.5 KH2, and the audio signal is band-limited by the filter 2, and this audio signal is supplied to the A/D converter 4 via the amplifier 3. .

For example, the A/D converter 4 has a sampling frequency of 12.
It is a 5KH2 8-bit A/D converter, and the audio signal is A/D converter.
/D converter 4 converts the signal into an 8-bit digital signal, which is then supplied to 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, the audio signal is expressed by the magnitude of the discrete frequency spectrum of N channels. Then, N-channel spectral data strings are output as one frame data every unit time (frame period). That is, the audio signal is extracted every frame period as a parameter expressed by an N-dimensional vector, and is supplied to the feature data extractor 6.

The feature data extractor 6 calculates the distance between adjacent frame data. For example, the absolute value of the difference in spectral data for each channel is determined, and the sum of the values is determined as the interframe distance.

Furthermore, the sum of the interframe distances is determined, and the trajectory length of the N-dimensional vector 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.

For example, the feature data extractor 6 extracts one frame data as shown in FIG.
-1) Frame data constituted by channel data is extracted to O frames (1-1) frames.

When this feature data is registered, it is stored in the registered pattern memory 7 as a registered feature data block. During recognition, the input audio signal undergoes the above-mentioned processing and is converted into an input feature data block, which is supplied to the pattern matching determiner 8, which performs pattern matching between the input feature data block and all registered data blocks. It will be done.

FIG. 3 shows an example of the pattern matching determiner 8,
As shown in FIG. 3, a frame distance calculation circuit 10, a weighting coefficient calculation circuit 11, a multiplication circuit 12, a matching distance calculation circuit 13, and a minimum distance determination circuit 14 constitute a pattern matching determiner 8.

The input feature data block is supplied from the feature data extractor 6 to the frame distance calculation circuit 10 and the weighting coefficient calculation circuit 11, and the registered feature data block to be compared is supplied from the registered pattern memory 7 to the frame distance calculation circuit 1.
0 and is supplied to the weighting coefficient calculation circuit 11.

In the frame distance calculation circuit IO, distance calculation between corresponding frames between the input feature data block and the registered feature data block is performed. The inter-frame distance data of the corresponding frames between the input feature data block and the registered feature data block obtained by the calculation process are multiplied by the multiplication circuit 1.
2.

Further, in the weighting coefficient calculation circuit 11, the difference between the spectral data of two adjacent channels is calculated for each spectral data string of the corresponding frame between the input feature data block and the registered feature data block, and the difference value is accumulated. calculated.

A weighting coefficient is calculated from the obtained cumulative difference value of the input waste vector data and the cumulative difference value of the registered spectrum data, and the weighting coefficient data is supplied to the multiplication circuit 12.

The multiplication circuit 12 multiplies the interframe distance data by the weighting coefficient data, and the interframe distance data multiplied by the weighting coefficient from the multiplication circuit 12 is supplied to the matching distance calculation circuit 13 as interframe distance data.

In the matching distance calculation circuit 13, the sequentially supplied interframe distance data is accumulated and the maximum frame (1-
When the interframe distance data up to 1) is accumulated, this accumulated value is used as matching distance data, and the matching distance data is supplied to the minimum distance determination circuit 14. Similarly,
Matching distances are calculated between all registered feature data blocks and input feature data blocks, and the matching distance data is supplied to the minimum distance determination circuit 14.

The minimum distance determination circuit 14 outputs, as a recognition result, a word corresponding to a registered feature data block for which the matching distance is the minimum and the distance is determined to be sufficiently close.

The operations of the frame distance calculation circuit 10, the weighting coefficient calculation circuit 11, and the multiplication circuit 12 of the pattern matching determiner 8 in the above-mentioned embodiment of the present invention will be explained with reference to a flowchart.

The registered feature data block is supplied from the registered pattern memory 7 to the frame distance calculation circuit 10 and the weighting coefficient calculation circuit 11, and the input feature data block is supplied from the feature data extractor 6 to the frame distance calculation circuit 10 and the weighting coefficient calculation circuit 11. 0 weighting coefficient calculation circuit 1□9 7
, each, ,'f:), Steps 1 to 2 are performed for each frame of m data blocks.

The weight Ws of the input waste vector data string composed of N channels from channel 0 to channel N-1 in the i-th frame is calculated as follows, where n is a variable indicating the channel number. The sum of the absolute values of the differences in the waste vector data is set as the weight Ws (step ■).

Next, the weight R3 of the registered spectrum data and the evening edition is similarly determined as follows, and the sum of the absolute values of the differences in the registered spectrum data between adjacent channels is set as the weight Wsl (step 2).

Then, in step (2), the weighting coefficient W5 is determined (step (2)). The process of step (2) is performed in the frame distance calculation circuit IO and the multiplication circuit 12.

In step (2), the i-th inter-frame distance Di is found as follows. That is, the absolute distance between corresponding channels of each frame is calculated, and a weighting factor W is added to the sum total from channel 0 to channel (N-1).

By multiplying by , the interframe distance is calculated.

Note that 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, the pattern matching determiner calculates the absolute value of the difference between the spectral data of adjacent channels for each spectral data string for each frame between the registered feature data block and the input feature data block, and calculates the absolute value of the difference between the spectral data of adjacent channels. The difference values are accumulated, the weight Ws of the input vector data string and the weight WjI of the registered spectrum data string are calculated, the weight coefficient Ws is calculated using the weight W3 and the weight W8, and the registered feature data block and the input feature are calculated. The interframe distance of the data block is multiplied by a weighting constant Ws, a matching distance is calculated by the sum of the multiplication results, and a matching judgment is performed based on the calculated matching distance.

For example, if there are three frames each composed of 10 spectrum data from channel 0 to channel 9 as shown in FIG. 5A, FIG. 5B, and FIG. 5C, each channel in FIG. The size of the vector data is (
10,14,12゜9. 6. 8. 10.9. 9
．． The weight WA of frame A shown in 8) is Wa ``4+2+3+3+2+2+1+O+1,
Also, the magnitude of the spectrum data of each channel in Fig. 5B is (10, 15, 12, 9°6.8, 15.8.7
．． 7) The weight Ws of frame B is Wa −5+3+3+3+2+7+'? +1+0, and the magnitude of the spectrum data of each channel in FIG.
The weight Wc of frame C shown in 9) is We-4+3+3+5+2+10+6+1+0, which is the weight for frame A and frame C, and the weight coefficient W3 for frame B and frame C is as follows. Frame open circuit ND between frame A and frame C
A is DA = 20X2.42-48.4, and the frame open circuit MDw between frame B and frame C is Ds
”20 x 2.01 = 40.2, frame C
It is determined that frame B is similar to the frame B.

As can be understood from the above example, according to the present invention, during pattern matching, the inter-frame distance is multiplied by a weighting coefficient corresponding to the amount of change in the spectrum in the frequency axis direction, so that the spectral data string is characteristics are strongly brought out. Therefore, even if many similar registered spectrum data sequences are registered in the registered pattern memory, the recognition rate does not decrease.

[Brief explanation of the drawing]

Fig. 1 is a block diagram showing the overall configuration of an embodiment of the present invention, Fig. 2 is a route diagram used to explain the data structure of the feature data block in an embodiment of the invention, and Fig. 3 is a block diagram showing the overall configuration of an embodiment of the invention. FIG. 4 is a block diagram of a pattern matching determiner in an embodiment of the present invention; FIG. Figures A, 5B, and 5C are route maps showing examples of spectrum data. Explanation of main symbols in the drawings: microphone, 5: acoustic analyzer, 6: feature data extractor, 7: registered pattern memory, 10: frame distance calculation circuit, 11: weighting coefficient calculation circuit, 12: multiplication circuit, 13 : Matching distance calculation circuit, 14: Minimum distance determination circuit.

Claims (1)

[Scope of Claims] Acoustic analysis means that performs preprocessing necessary for speech recognition, such as spectrum conversion, on an input audio signal; and feature data extraction that is supplied with output data of the acoustic analysis means and extracts feature data from the output data. means, a memory in which the feature data is stored as a standard pattern, an input pattern from the feature data extraction means, and the standard pattern read from the memory;
Each spectrum data S forming the above input pattern
A value W_s corresponding to the amount of change in the frequency axis direction of _i_n and each spectrum data R forming the above standard pattern.
The weighting coefficient W_i is calculated using the value W_R corresponding to the amount of change in the frequency axis direction of _i_n, and the distance between the input pattern and the standard pattern is calculated by multiplying the inter-frame distance by the weighting coefficient W_i. A speech recognition device comprising: pattern matching determining means for determining matching based on processing results.