JPS62105199A - 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] This invention relates to a speech recognition device used for recognizing the speech of a specific speaker word by word, for example.

Invention C (Already Needed) This invention performs time axis normalization processing using an adaptive number of divisions that is approximately proportional to the length of the trajectory of a multidimensional vector expressed by, for example, a 749-dimensional 1ml Helix number spectrum. In the pattern matching judger of the speech recognition device that selects the means for extracting data, performs a swing search, selects the minimum interframe distance, and calculates the selected interframe distance ('*i13). By calculating the distance between the input pattern and the standard pattern while shifting the center value of the search width accordingly, it is possible to avoid cases where the number of frames that make up the input pattern and the standard pattern are slightly different, or where there is a change in the time axis of the audio. This is to ensure accurate matching even when the pattern after time axis normalization is largely influenced by time axis fluctuations.

[Conventional technology]

A conventional voice recognition WiF is configured with a microphone as a voice input unit, 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 sent to the preprocessing circuit.
However, the bandwidth is limited to that required for speech recognition.
A D 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 cycle). That is, data of one frame every K months (frame 1'qI) is taken as frequency spectrum data of the X channel, cut out as a parameter expressed by an N-dimensional vector, and supplied to the feature data extraction section.

In the feature data extractor, the distance between adjacent frame data is calculated. The absolute value of the difference between the spectral data of the corresponding channels is 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). During recognition, the input audio signal undergoes the above-mentioned processing, becomes an input data block (input pattern), and is supplied to the pattern matching determiner, which compares the input feature data block with the registered feature data block. Pattern matching is performed.

In the pattern matching determiner, a matching distance between the input feature data block and the registered feature data block is calculated. For example, one frame data is extracted by the feature data extractor, and the feature data pro- gram is composed of frames 0 to N-1). At the time of registration, between the frame data that makes up the 1 revised data program and the frame data that makes up the input feature data block,
While the swing search is being performed, the distance between corresponding frames is calculated.

For example, for the i-th frame interrogation^1tsDi, n is the channel number, input waiting 1"5!i, data block spectrum data is 81., registered feature data pro,
The spectrum data R1, . That is, the first frame of the input waiting data block, the (i-1)th frame of the registered feature data block, the i-th frame of the input waiting data block, and the first frame of the special data block. , Input waiting (1) The sum of the absolute values of the differences of the corresponding channels between the i-th frame of the data block and the (i+1)-th frame of the e ff & feature data block is calculated, and the obtained three values are calculated. The minimum value among them is taken as the interframe distance SD.

Then, since swing search is not performed for all frames of the input feature data block, registration 1. The interframe distance SD between the frame of the specific ft data block is determined, and the sum of this interframe distance SD up to (i=I-1), that is, the matching distance is determined. The matching distance is similarly calculated for the registered feature data block of the pond, and the Chinese language corresponding to the registered feature data block for which the matching distance is the minimum and is determined to be sufficiently close is output as a recognition result.

In addition, while performing a swing search, the pattern matching determiner determines a weighting coefficient that is approximately proportional to the search width and approximately inversely proportional to the number of frames constituting the input feature data block. Japanese Patent Application No. 150697/1983 has been proposed by the applicant of the present invention as a voice recognition device that performs distance calculation processing by multiplying the distance between the two by this weighting factor.

[Problem that the invention seeks to solve]

Using a swing search in a conventional pattern matching judger, the inter-frame distance is determined by the input waiting t''sl data according to the search width (\r-m), as described above. (i - m) frame ~ (i + m) of the feature data block registered in the first frame of the block
) is determined by matching the frames up to the frame.

However, for example, in a special tTh data extractor, the number of divisions is variable depending on the length of the trajectory of a multidimensional hector expressed by a discrete frequency spectrum, and feature data is extracted depending on the word length of the speech. In the method in which the time axis is normalized, the number of frames of the registered data block is registered as being various.

For this reason, it is possible to search for patterns using the conventional swing search between the registered feature data block, which has a slightly different number of frames than the input waiting data block, and the number of frames differs by more than the search width, and the input waiting ifi data block. When matching is performed, accurate matching can be performed.L'7
There was a problem that the 141j rate decreased.

In addition, if the time axis fluctuations of the audio are large and the effects of the time axis fluctuations remain in the feature data block even after time axis normalization in the feature data extractor, matching cannot be performed accurately and recognition cannot be performed. , there was a problem that the comprehension rate decreased.

Therefore, an object of the present invention is to perform accurate matching even when the number of frames of a large-scale feature data block and a registered feature data block are slightly different, or when the influence of time @ fluctuation remains in the feature data block. The purpose of the present invention is to provide a voice recognition/claw device that has a means to enable this.

[Means for solving problems]

In this invention, an acoustic analysis means 5 performs preprocessing necessary for speech recognition such as spectral modification (Q) of an input speech signal, and the output data of the acoustic analysis means 5 is supplied, and the time axis is normalized by an appropriate number of divisions. a feature data extraction means 6 that performs processing to extract feature data; a memory 8 in which the special ITh data is stored as a single small pattern; The found target Y 廿 pattern is supplied, and a frame constituting one of the input pattern and the speciation pattern, and a plurality of frames constituting the other of the +111 Ft input pattern and the facial expression pattern corresponding to the search width. The distance between each frame is determined by corresponding the distance between the frames, and the center of the search width is shifted in accordance with the input information of the minimum interframe distance, and the distance between the 5 input patterns and the pattern is calculated. , a pattern matching determination means 9 that performs a matching determination based on the result of distance calculation processing;

[Effect]

In the pattern matching determiner 9, the frames of the registered feature data block are made to correspond to the frames of the input feature data block according to the search width, and the interframe distance between each frame is determined, and the At the same time, it is determined which frame of the registered feature data block is used to calculate the minimum inter-frame distance. If the minimum inter-frame distance tends to be calculated, the center value of the search width is shifted in the middle direction, and if the minimum inter-frame distance tends to be calculated from frames in one direction. In this case, the center value of the search width is shifted in one direction, inter-frame distances are calculated sequentially, and the minimum inter-frame distance is accumulated to calculate the man-amp 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, and in FIG. 1, 1 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.

To-off frequency 7. ．． This is a low-pass filter of 5KN, , and the voice signal is limited to the band below 7,5H2 required for voice recognition in the filter 2, and this voice signal is passed through the amplifier 3 to the A/D. It is supplied to converter 4.

For example, the A/D converter 4 has a sampling frequency of 12.
5K) I2 is an 8-bit A/D converter, and the audio signal is analog-to-digital converted into an 8-bit digital signal in the A/D converter 4, which is then converted into an 8-bit digital signal.
supplied to

The acoustic analyzer 5 detects the audio signal as 1? By converting it into an i1 wave number spectrum, a spectral data string of, for example, N chances is generated. 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 the X channel. And unit time (frame period!t
JI), N-channel spectral data strings are output as one frame data. That is, the audio signal is extracted for each 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. The absolute value of the difference between the spectral data of the corresponding chamble is 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, the trajectory length is divided by a number of divisions approximately proportional to this trajectory length, and only the frame data corresponding to the division point is extracted as feature data. In other words, the time axis is normalized so as not to be affected by fluctuations in the rate at which the speaker's voice is generated, and a number of frame data that exhibit strong unique characteristics are extracted and output in accordance with the word length of the voice. .

When this characteristic data is registered, the mode switching circuit 7
The data is supplied to the registered pattern memory 8 via the registered pattern memory 8 and stored as registered feature data.

At the time of recognition, the input audio signal undergoes the processing described above and is made into feature data, which is supplied to the pattern matching determiner 9 and is made into an i-sheet data block waiting for input.

In the pattern matching judger 9, only the 62+= #* ffi data block whose number of frames is not significantly different from the number of frames of the input waiting FJ data block is compared, and the input waiting FJ data block and the registered feature data block are compared. In between, pattern matching is performed.

FIG. 2 shows an example of the pattern matching determiner 9,
As shown in Fig. 2, the search width frame distance calculation circuit 1
0. Minimum value extraction circuit 11. Search width interval variable calculation circuit 1
2. Maching distance calculation circuit 13. Addition circuit 14. End-side frame distance calculation circuit 15. A pattern matching determiner 8 is configured by the end-side matching distance calculating circuit 16 and the minimum distance determining circuit 17.

The feature data from the feature data extractor 6 is supplied to each of the frame distance within search width calculation circuit 10 and the end-side frame distance calculation circuit 15, and, for example, the frame corresponding to the end of the voice section is extracted (Q-1). In this case, frame data each composed of data from channel O to channel N-1 is supplied from frame 0 to frame Q-1 and used as an input feature data block. At the same time, a registered feature data block composed of, for example, O frame width - 1 frame, which is not significantly different from the frame number of the input feature data block to be compared from the registered pattern memory 8, has a frame distance within the search width. The signal is supplied to the calculation circuit 10 and the end-side frame distance calculation circuit 15.

In the search width frame distance calculation circuit 10, when searching from the input feature data block side to the registered feature data block side, q frames on the input feature data block side are j frames and j on the registered feature data block side.
A plurality of frames in the vicinity of the frame are matched, and an interframe distance between each of them is calculated.

For example, when the corresponding frame width, that is, the search width is set to 1, as shown in FIG.
-1) Frame, J frame, and (j+1> frame are matched, and the interframe distance between them is calculated. Also, when the search width is set to 2, it is registered in the q frame of the input feature data block. Feature data block (j
-2) frames to (j+2) frames are matched, and the interframe distance between each of them is calculated. Furthermore, when the search width is set to m, q frames of the input feature data block are closely matched with (j-m) frames to (j 10 m) frames of the registered feature data block, and (2m=1) frames are Inter-frame distances are calculated respectively.

(2m+1) pieces of interframe distance data obtained through the calculation process are supplied to the minimum value extraction circuit 11.

In the minimum value extraction circuit 11, each of the (2m=1) pieces of interframe distance data is compared, and the interframe distance data having the minimum value is selected as the interframe distance for q frames of the input feature data block. The interframe distance data is supplied to a matching distance calculation circuit 13. At the same time, frame number data indicating which frame in the registered feature data block the selected interframe distance data was obtained is supplied to the search width interval variable calculation circuit 12.

In the search width section variable calculation circuit 12, the shift direction of the search width section is determined based on the frame number data, and shift operation data is generated.

For example, based on frame j of the first attack data block at registration, which is the center of the search width section corresponding to frame q of the input feature data block, frames between frame j - (j+m) tend to be selected. If it is determined that this is the case, shift operation data for shifting the search width section in the ten-direction direction by j frames - (j''m) frames is generated.

In addition, if it is determined that frames between (j-m) frame number data tend to be selected, one-way,
That is, shift operation data for shifting the search width section from frame (j-m) to frame j is generated.

This shift operation data is supplied to the search width frame distance calculation circuit io, and the inter-frame distance calculation for each frame of the human-powered feature data block is sequentially performed on the registration time i attack data block l set by the shift operation data.
This is done between frames within the search width section of jl, +1. In other words, the search width section is shifted so that the minimum interframe distance is always calculated from the frames within the search width section of the registered feature data block, and (2m+1) interframe distances are calculated. The interframe distance data selected as the minimum value in the minimum value extraction circuit 11 is supplied to the matching distance calculation circuit 13.

In the matching distance calculation circuit 13, the interframe distance data sequentially supplied and selected as the minimum value is accumulated, and the maximum frame (
When the inter-frame distance data is accumulated up to Q-1), this accumulated value is used as matching distance data, and the matching distance data is supplied to the adder circuit 14.

Further, in the end-side frame distance calculation circuit 15, an inter-frame distance is calculated for a predetermined end-side frame of the input feature data block and a predetermined end-side frame of the registered feature data block. For example, when the interframe distance is calculated by direct matching, and C is a constant number of frames, registration is made in each frame from the predetermined frame (Q-C) to the maximum frame (Q-1) of the input feature data block. A given frame of feature data block (
JC”) to the maximum frame (J-1) are each 1
The interframe distances are sequentially determined, and the interframe distance data is sent to the terminal side matching distance calculation circuit 1.
6.

In the end-side matching distance calculation circuit 16, the sequentially supplied inter-frame distance data is accumulated, and when the inter-frame distance data is accumulated up to the maximum frame (Q-1) of the input feature data block, this accumulated distance data is accumulated. The calculated value is used as matching distance data obtained for the end frame, and the matching distance data is supplied to the adding circuit 14.

In the adding circuit 14, the matching distance data supplied from the matching distance calculation circuit 13 and the matching distance data supplied from the end-side matching distance calculation circuit 16 are added to form determination distance data, and this determination distance data is The signal is supplied to the minimum distance determination circuit 17. Also,
Similarly, the determination distance is calculated between other registered feature data blocks with a large number of frames (not different),
The distance data for determination is supplied to the minimum distance determination circuit 17.

The minimum distance determination circuit 17 outputs, as a recognition result, a word corresponding to a registered feature data block that is determined to have a minimum determination distance and is sufficiently close.

Fig. 4 shows the change in the search width section when, for example, the search width is set to 1 and frames in the +1 direction or -1 direction of the registered feature data block are selected twice or more in succession when extracting the minimum value of the interframe distance. It shows an example of a process in which an operation is performed and a determination distance is calculated.

In FIG. 4, the vertical axis indicates the frame number of the registered feature data block, and the horizontal axis indicates the frame number of the input feature cheater block. Each point on the thin line in Fig. 4 indicates the frame number corresponding to the search width section that has been variably manipulated, and each point on the broken line in Fig. 4 indicates the frame number when the search width section has not been variably manipulated. , each point on the solid line in FIG. 4 indicates a frame number adopted as the determination distance.

As shown in FIG. 4, only when the +1 direction or -1 direction of the registered feature data block is selected twice in succession, the search width section is shifted, the interframe distance is calculated, and the input waiting fl data block is A matching distance between the registered feature data block and the registered feature data block is determined. Further, the interframe distance is calculated only for the predetermined terminal side frames of the input feature data block and the registered feature data block, and V! The matching distance for the end measurement frame is determined.

Then, the matching distance determined for the entire frame and the matching distance determined only for the end frame are added to form a determination distance, and this determination distance is used to determine word recognition and identification.

Frame distance calculation circuit 10 within the search width of the pattern matching determiner 9 in the embodiment of the present invention described above. Minimum value extraction circuit 11. The operations of the search width interval variable calculation circuit 12 and the matching distance calculation circuit 13 will be explained with reference to the flowchart shown in FIG. 5, for example, with the search width set at one. In addition, the flowchart in Fig. 5 shows that when the 4-1 direction or the 11 direction of the frame of the registered feature data block is selected twice or more consecutively when extracting the minimum value of the inter-frame distance, the search width section variable operation is performed. This applies when

In the search range frame distance calculation circuit 10, the processes of steps (2) and (2) are performed between the input feature data block and the registered feature data processor.

First, the number of times K of selections in one direction between frames of the registered feature data block and the number of times P of selections in one direction of each frame of the registered feature data block are initialized to 0, and a variable indicating the frame number of the registered feature data block is initialized. j and the variable q indicating the frame number of the input feature data block are 0.
is initialized to . Additionally, the cumulative distance between frames determined by the swing search is initially set to O, the maximum frame number of the registered feature data block is set to J-1, and the maximum frame number I of the input feature data block is set to O. Q-1, and the limit value Δ for the search width section is set to a predetermined value (step ■).

Then, in step (2), the q frames of the input feature data block are associated with the (j+1), j, and (j −1) frames of the registered feature data block, and the interframe distances between them are D I+ D z,
D:l is determined by using n as the channel number and SQn as the input cassette vector data+Rjn as the registered spectrum data. That is, the absolute value of the difference between the spectral data of the corresponding channels from channel O to channel (N-
The sum up to 1) is the interframe distance D+, D z, D
1. In addition, in the process of step (8), when Cj of the input feature data block is set to O), since (j −N frame does not exist in the i-attack data block side at the time of registration, (j −1− 〇).Also,
If the number of (j+1) frames is greater than or equal to the maximum frame number of the registered feature data block, there is no (j+1) frame in the registered feature data block, so
=L).

Minimum value extraction circuit 11. In the matching distance calculation circuit 13 and the search width section variable calculation circuit 12, steps
Processing of O is performed. In step ■, the interframe distance AID obtained by associating (j↓1) frames with q frames of the input feature data block, and the frame dark path obtained by associating j frames with q frames of the input feature data block. ^ID2 and the frame dark path Ai D z obtained by associating the (j-1) frame with the q frame of the input feature data block are compared, and the smallest one is the interframe for the q frame of the input feature data block. considered to be the distance. This inter-frame distance is accumulated to form a cumulative distance.

In addition, in step ■ and step ■, the interframe distance selected as the minimum is registered ((j
+1) frame, j frame, and (j-1) frame is determined.

In step ■, it is determined whether or not the interframe distance is selected, and if DI is selected,
The number of times K of direction selection is incremented, and -1
The number of times P of direction selection is set to 0 (step ■). Also,
If DI is not selected, it is determined in step (a) whether or not the interframe distance D3 is selected, and D is determined.
If 3 is selected, the number of selections P in the -1 direction is incremented, and the number of selections K in the -1 direction is set to 0 (step 0). Further, if Dl and D3 are not selected, that is, if D2 is selected, step
In this case, the number of selections K in the +1 direction and the number P of selections in the -1 direction are both set to 0.

In steps (2) and (2), it is determined whether the search width section should be varied.

In step ■, the number of selections in the +1 direction is compared with 2 to determine whether frames in the +1 direction have been selected twice or more in a row, and if (K≧2), the registered feature It is determined whether the difference (j-q) between the variable j indicating the frame number of the data block and the variable q indicating the frame number of the input amnesty cheater block is less than or equal to the limit value Δ (step 0). Then, 1rll l (if the old direct Δ is less than or equal to Δ, j is incremented and set to 1 minus 1 (
Stenobu D), if the limit value Δ is exceeded, j is left as is and the process moves to the next step.

Moreover, in the case of (K<2), in step (■), -
By comparing the number of selections P in one direction with 2, it is determined whether frames in one direction have been selected twice or more in a row, and if (P≧2), the input waiting data is Difference between variable q indicating the frame number of the block and variable j indicating the frame number of the registered feature data block (q-j)
It is determined whether or not is less than or equal to the limit value Δ (step @).

Then, if it is less than the limit value Δ, j is decremented to j-1 (step O), and if it exceeds the limit value Δ, j is left as is and the process moves to the next step.

In step ■, the variable q indicating the frame number of the input feature data block is compared with the maximum frame number ■ of the input feature data block, and if (q<1), the variable q indicating the frame number of the registered feature data block is At the same time, j is incremented to j↑1 (step [phase]), q is incremented to q''1, and the processes of steps ■ and ■ are performed again. Variable q indicating the frame number of the input feature data block and variable j indicating the frame number of the registered feature data block are Both are incremented by 1, and the calculation process for the next frame is started.

Additionally, if frames in the +1 direction are selected twice or more consecutively, q is incremented by +1, and j is incremented by +/-.
Since it is incremented by 2, the search width section is shifted in the +1 direction, and the calculation process for the next frame is started. - If frames in one direction are selected twice or more in succession, only q is incremented by 10l, so -1
The search width section is shifted in the direction, and calculation processing for the next frame is started.

By repeating the above process, the interframe distance is calculated for each frame while the search width section is shifted, the minimum interframe distance is accumulated, and the variable q indicating the frame number of the input feature data block is calculated. is set as the maximum frame number ■, and when the calculation process is performed on the end frame, the calculation on one registered feature data block is completed. The cumulative distance AID of the inter-frame distances at this time is taken as the matching distance. Then, matching distances are similarly calculated during other registrations (between offensive data blocks).

In one embodiment of the present invention, a case has been described in which the search is performed from the input feature data block side to the registered feature data block side, but it is also possible to perform a search from the registered feature data block side to the input feature data block side. good. In addition, in one embodiment of the present invention, when determining the distance between frames on the end side, a case has been described in which the distance between frames is determined by direct matching. It may also be configured to calculate the distance. Also,
In one embodiment of the present invention, a case has been described in which the variable operation of the search width section is performed when frames in the +1 direction or -1 direction of the registration waiting data block are selected twice or more in succession. The present invention is not limited to the above, and a configuration may also be adopted in which variable operations are performed based on probabilistic judgments.

Moreover, this invention is not limited to a hard-wired configuration.
The processing may be performed by software using a microcomputer or a microprogram method.

〔Effect of the invention〕

In this invention, in the pattern matching judger, a plurality of frames of the registered feature data block are made to correspond to a frame of the input waiting data block according to the search width. is calculated, the minimum inter-frame distance is selected, it is determined which frame the minimum inter-frame distance was calculated by the registered feature data professional, and the tenth of the registered feature data blocks are determined. If the minimum inter-frame distance tends to be calculated from the frames in one direction, the center value of the search width is shifted in the direction, and the minimum inter-frame distance is calculated from the frames in one direction. If there is a tendency, the center value of the search width is shifted in one direction so that the minimum inter-frame distance is always calculated within the search width, and the inter-frame distances are sequentially calculated and the minimum distance is calculated. The matching distance is calculated by accumulating the inter-frame distances.

Therefore, according to the present invention, it is possible to cope with the case where the number of frames of the input feature data block and the registered feature data block are slightly different, or even when the sound with time axis fluctuation remains in the feature data block. It is possible to perform accurate matching and improve the recognition rate.

Further, in one embodiment of the present invention, a partial matching distance between a predetermined end frame of the input feature data block and a predetermined end frame of the registered feature data block is calculated, and this matching distance is Since it is added to the overall matching distance obtained while shifting the search width between the attack data block and the registered feature data block and used as the judgment distance, it is possible to strengthen the difference in the endings of the partner words, and The recognition rate can be improved.

[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 block diagram of a pattern matching determiner in an embodiment of the invention, and FIGS. 3 and 4 are diagrams of an embodiment of the invention. FIG. 5 is a route map used to explain the operation of the pattern matching determiner in the embodiment, and FIG. 5 is a flowchart used to explain the operation of the pattern matching determiner in the embodiment of the present invention. Explanation of main symbols in the drawings 1: Microphone, 5: Acoustic analyzer. 6: Feature data extractor, 7: Mode switching circuit. 8: Registered pattern memory, 9: Pattern matching determiner, 10: Frame distance calculation circuit within search width,
11: Minimum value extraction circuit, 12: Search width section variable calculation circuit, 13: Masoching distance calculation circuit,
14: Addition circuit, 15: Termination side frame distance calculation circuit, 16: Termination side matching distance calculation circuit, 1
7: Minimum distance determination circuit. Agent Patent Attorney Tadashi Sugiura Tomokai 1A Figure 1 Figure 2

Claims (1)

[Claims] Acoustic analysis means that performs preprocessing necessary for speech recognition such as spectrum conversion of an input audio signal; and output data of the acoustic analysis means is supplied, and the time axis is normalized by an appropriate number of divisions. a feature data extracting means for processing and extracting feature data; a memory in which the feature data is stored as a standard pattern; and an input pattern from the feature data extracting means and the standard pattern read from the memory. The distance between each frame is determined by associating the supplied frame constituting one of the input pattern and the standard pattern with a number of frames corresponding to the search width constituting the other of the input pattern and the standard pattern. is calculated, and the distance calculation process between the input pattern and the standard pattern is performed while shifting the center value of the search width according to the information on the minimum inter-frame distance, and matching is determined based on the result of the distance calculation process. A speech recognition device comprising: a pattern matching determination means;