JPH0277100A - Preliminary selecting device for speech recognition - Google Patents

Abstract

PURPOSE:To make an effective preliminary selection by deleting distance data corresponding to all feature patterns of a group that a feature pattern for preliminary selection which is selected and outputted belongs to, from a distance data storage means. CONSTITUTION:A preliminary selecting means 4 selects and outputs the feature pattern for preliminary selection which has the shortest distance among distances stored in a distance data temporary storage means 3. A link data storage means 5 is stored previously with the relation between respective feature patterns and all feature patterns in the group that the feature pattern belongs to. Then all the feature patterns of the group that the feature pattern for preliminary selection which is selected and outputted belongs to is outputted from the link data storage means 5 and inputted to the distance data storage means 3, thereby deleting the distance data corresponding to all the feature patterns of the group that the feature pattern for preliminary selection belongs to from the distance data storage means 3. Consequently, a single feature pattern for preliminary selection is obtained for the contents of one input voice and the width of the preliminary selection is never narrowed down.

Description

[Detailed Description of the Invention] [Summary] Regarding the improvement of a preliminary selection device for speech recognition that preliminarily selects a small number of feature patterns from a speech dictionary having a plurality of feature patterns for one input speech content during speech recognition, The purpose of the present invention is to provide a preliminary selection device capable of effective preliminary selection without narrowing the range of preliminary selection, and a preliminary selection device that registers and stores in advance a group of a plurality of preliminary selection feature patterns corresponding to each human voice content. a distance calculation means for calculating a distance between an input feature pattern created from the input audio content and each preliminary selection feature pattern stored in the preliminary selection feature pattern storage means; a distance data temporary storage means for temporarily storing distance data corresponding to the distance data temporary storage means; and a pre-OMM that selects and outputs a preliminary selection feature pattern indicating the minimum distance among the distances stored in the distance data temporary storage means. link data storage means for storing in advance the relationship between each feature pattern and all feature patterns of the group to which the feature pattern belongs; A feature pattern is outputted from the link data storage means and inputted to the distance data storage means, so that the preliminary selection feature pattern selectively outputted from the distance data storage means corresponds to all feature patterns of the group to which it belongs. Configure to delete distance data.

[Industrial application field]

The present invention relates to an improvement in a preliminary selection device that preliminarily selects a small number of feature patterns from a speech dictionary having a plurality of feature patterns for one human voice content during speech recognition.

In recent years, with advances in speech recognition processing technology and advances in LSI technology, many speech recognition devices have been developed, and applications such as use in man-machine interfaces are being considered.

Since the voice recognition device is used as a man-machine interface, the bi needs to meet requirements such as small size, low cost, and high performance. In particular, even when the sound of words that express the same content is pronounced, there may be fluctuations in each utterance, differences in voice depending on the speaker, mood at the time of pronunciation, etc.
Depending on the environment, time factors such as utterance time and rhythm, level factors such as voice volume and accent, and frequency component factors such as changes in intonation and formant position may vary, but are always constant and stable. Therefore, it is necessary to avoid misrecognition even in such cases.

Therefore, for the purpose of improving performance, a method has been proposed in which a plurality of voices are used during voice registration, and characteristic patterns are created and registered for each voice (multi-template registration method). In other words, for example "1" with the same content, [Ichi J
Characteristic patterns are created and registered for each case, such as a short pronunciation of ``,'' and a long pronunciation of ``each''. This has the effect of reducing the possibility of erroneous recognition due to variations in speech rhythm due to variations in input speech during recognition.

However, in this registration method, multiple feature patterns are created for one human voice content, so the number of feature patterns registered in the speech dictionary increases, and as a result, during recognition, the features created by the input speech The amount of distance calculation between the pattern and the feature pattern based on the speech dictionary increases. In order to provide a good response, it is necessary to perform calculations at high speed, and increasing the amount of distance calculations as described above not only goes against this necessity, but also runs the risk of going against the demands for small size and low cost. There is.

One method to solve this problem is to obtain a simple distance using only specific parameters from the feature patterns stored in the speech dictionary, perform preliminary selection using this simple distance, and select the preselected candidates. Among these methods, there is a preliminary selection method that performs highly accurate matching.

FIG. 4 is a diagram showing the system configuration of a speech recognition device including a preliminary selection method which is the background of the present invention. The present invention relates to an improvement of the preliminary selection processing section 405 in the system configuration shown in FIG.

[Prior Art] FIG. 5 is a block diagram showing an example of a conventional preliminary selection device. In the same figure, according to the conventional preliminary selection method, specific parameters of the input voice feature pattern obtained by the manual pattern register 501 and all the feature patterns in the speech dictionary memory 502 are used for preliminary selection. A simple distance calculation unit 503 calculates a simple distance, and a distance data temporary stack memory 505 and a minimum distance determination unit 50
6. The counter 507 selected the top N feature patterns in descending order of the simple distance. The registered pattern number data memory 504 stores the pattern number in the dictionary memory 502.
Output to. The pattern number obtained from the output of the minimum distance determination unit 506 is stored in the distance data temporary stand memory 505.
, and the pattern number is deleted from the distance data temporary stack memory 505. When the counter 507 counts the pattern number output from the minimum distance determination section 506 N times, the counter 507 stops the output from the minimum distance determination section 506. The output from the minimum distance determination unit 506 is stored in the memory 508 as a preliminary (!li! selection result).

However, when the speech dictionary memory 502 has a plurality of feature patterns for one input speech content as described above, if the conventional preliminary selection method is applied as is, it will not be possible to predict the plurality of feature patterns for the same speech content. M may end up being selected.

[Problem to be Solved by the Invention] Therefore, in such a case, even though N feature patterns have been preliminarily selected, N or less pre-selections are made for the content of the registered voice, and the preliminary selection is ! This caused problems such as the narrow area of the river.

SUMMARY OF THE INVENTION An object of the present invention is to solve such problems and provide a preliminary selection device capable of effective preliminary selection without narrowing the range of the preliminary Jx amount.

[Means to solve the problem]

FIG. 1 is a block diagram of the principle of the present invention. In the figure, the preliminary selection device for speech recognition according to the present invention includes a preliminary selection feature pattern storage means 1, a distance calculation means 2, a distance data temporary storage means 3, a preliminary selection means 4, and a link data storage means 5. and (l!ff).

The preliminary selection feature pattern storage means 1 registers and stores in advance a group of a plurality of child 61 U tR feature patterns corresponding to each input voice content.

The distance calculating means 2 calculates the distance ^1 between the human feature pattern created from the input voice content and each preliminary selection feature pattern stored in the preliminary selection feature pattern storage means 1.

The distance data temporary storage means 3 temporarily stores distance data corresponding to each preliminary selection feature pattern.

The preliminary selection means 4 selectively outputs a preliminary (ii# selection feature pattern) indicating the minimum distance behind the distance stored in the distance data temporary storage means 3.

The link data storage means 5 stores in advance the relationships between each feature pattern and all feature patterns of the group to which the feature pattern belongs.

All feature patterns of the group to which the selectively output preliminary selection feature pattern belongs are outputted from the link data storage means 5 and inputted to the distance data storage means 3. The distance data corresponding to all feature patterns of the group to which the selection feature pattern belongs is deleted.

(Operation) The group to which the preliminary selection feature pattern selectively output from the preliminary selection means 4 belongs is specified by the link data storage means 5, and the distance data corresponding to the specified feature pattern is stored in the distance data temporary storage means 3. As a result, the output of the preliminary selection means 4 is a single preliminary selection feature pattern for one input audio content, instead of a plurality of characteristic patterns being preliminary selected as in the past. , so the range of preliminary selection is not narrowed.

〔Example〕

FIG. 2 is a block diagram illustrating a preselection device according to an embodiment of the present invention. In this figure, the same parts as in FIG. 1 are given the same reference numerals. That is, the preliminary selection feature pattern storage means 1 includes a registered pattern number data memory 202 and a dictionary memory 203, and the distance calculation means 2 includes a preliminary selection simple distance calculation section 204, which stores distance data. The temporary storage means 3 is composed of a distance data temporary stack memory 205, and the preliminary selection means 4 is composed of a minimum distance determination section 206. The link data storage means 5 includes a pattern number-voice number link data memo 'J209, a collation section 210, and a re-verification section 211.

Dictionary memory 203 includes a plurality of characteristic patterns and pattern numbers corresponding thereto.

The pattern number-voice number link data memory 209 stores link data that links pattern numbers and voice numbers.

Corresponding to one human feature pattern obtained from the output of the input pattern register 201 by unknown input speech, N features are selected from among the feature patterns stored in the dictionary memory 203.
When preliminarily selecting a sheet pattern, the preliminary fJM selection simple distance calculation unit 204 uses the input feature pattern and the dictionary memory 203.
Calculate the distance from the feature pattern corresponding to all pattern numbers in . The distance data temporary storage memory 205 temporarily stores the data calculated by the preliminary selection simple distance calculation unit 204 together with the pattern number of the feature pattern used in the calculation. The minimum distance determining unit 206 searches for distance data indicating the smallest value from among the distance data stored in the distance data temporary stack memory 205, and outputs a pattern number corresponding to this distance data.

In the link data storage means 5, the collation unit 210
The voice number corresponding to the pattern number output from the minimum distance determination unit 206 is retrieved from the pattern number-voice number link data memory 209, and the voice number is retrieved from the re-verification unit 211.
Enter. The re-verification unit 211 verifies one or more pattern numbers against the input voice number and outputs the pattern number.

The distance data temporary stack memory 205 deletes the data corresponding to the pattern number output from the re-verification unit 211 from the stored distance data.

When the counter 208 counts N pattern numbers, it causes the minimum distance determination unit 206 to stop outputting. N
The pattern numbers are sequentially stored in the preliminary selection result memory 207.

In this way, the N pattern numbers output from the minimum distance determining section 206 will not correspond to the same voice number.

As described above, in the embodiment of the present invention, the pattern number-voice number link data is used at the time of preliminary selection, and when performing preliminary selection of N feature patterns, the matching unit uses the pattern number output from the minimum distance determination unit 206. 210 and the re-verification unit 211, pattern numbers having the same voice number are detected, and distance data temporary stack memory 205 corresponding to the pattern number output from the re-verification unit 211 is used.
Delete the data stored in.

Therefore, the pattern number output from the minimum distance determining section 206 will not have the same voice number as the pattern number output before outputting it.

FIG. 3 shows the book selection processing unit 40 in the system shown in FIG.
7 is a block diagram showing the configuration of FIG. In the figure, the main selection processing unit 407 has an input pattern register 301,
It includes a dictionary memory 302, a main selection distance calculation section 303, a minimum distance calculation section 304, a pattern number-voice number link data memory 305, and a collation section 306.

The human pattern register 301 is the input pattern register 20
1 and outputs one feature pattern based on one human voice. The dictionary memory 302 outputs N preliminary selected feature patterns based on the pattern numbers output from the preliminary selection result memory 207 shown in FIG. The main selection distance calculation unit 303 calculates the distance between the feature pattern from the input pattern register 301 and the feature pattern from the dictionary memory 302. The minimum distance calculation unit 304 determines the minimum distance among the distances, and outputs a pattern number corresponding to the minimum distance. The matching unit 306 reads out the voice number corresponding to the input pattern number from the pattern number-voice number link data memory 305 and outputs it.

In the system configuration of the speech recognition device shown in Fig. 4, the one shown in Fig. 2 is used for the preliminary selection processing section 405, and the one shown in Fig. 3 is used for the main selection processing section 407, and the system is applied to 10-digit speech recognition. The operation of the embodiment when an automatic dialing device is constructed will be described below. Each block will be explained below.

Shigo U1 Tadashi A4- Perform band division using a 12-channel band pass filter (BPF), calculate the rectified (absolute value is taken, and smoothed by LPF) value for each channel, and io
The value of each channel is determined every ms.

The 12 channels' worth of output output for each IoffIs is logarithmically transformed, then the average of the 12 channels' worth of data is determined, and the difference between the average value and the value of each channel is output.

1 x 1. Section 403 Detects the start and end of a voice section by determining whether the value of human voice power is greater or less than a certain dark value level,
The data output from the feature extraction unit 402 in that section is stored.

(Method of Detecting the Starting Edge) When the input audio power is higher than the threshold value for 5 consecutive frames (50 ms) or more, the starting edge is defined as the point in time when the power changes from low to high.

(Method for detecting the end) After detecting the start end, when the human voice power is smaller than the dark value for 30 consecutive frames (300 ms) or more, the time point when the power changes from large to small is defined as the end.

■, 11 Yu no Seidai 619 parts 404 Voice sections of different lengths from the beginning to the end are 8 on the time axis.
It is divided and averaged in the divided sections. This results in 12 (channels) x 8 (frames) - 9
6' (parameter) characteristic pattern is obtained.

144 - Using learning audio, with three different levels of darkness values,
For one speech input, three characteristic patterns are created and recorded using the speech section detection section and normalization processing section.

The audio content to be registered is “1°” for audio number 1.
For voice number 2, it is "2°', etc., and all 10 numeric words are registered. However, for voice number 10, it is "2°', etc.
Let it be zero°′.

In other words, the number of registered feature patterns is 30.

゛ 几 B 405 This is an application part of the present invention, and its configuration is as shown in FIG.

The feature pattern in this example is 12 channels x 8 frames - 96 parameters, and the distance for pre-i selection is even channels (2.4°6, 8.10.1
In 2), for each frame, the absolute value of the difference between the parameters corresponding to the manually generated feature pattern and the feature pattern in the dictionary is determined, and the sum of these values is used.

Further, the number N of selection patterns for preliminary selection is assumed to be 10 patterns.

Therefore, by making a preliminary selection from the 30 patterns in the dictionary, a new IO pattern dictionary is created.

It's off and I'm depressed - child wli! One pattern (one voice) is selected from among the 10 patterns created by the selection processing unit 405. Its configuration is as shown in FIG.

The distance calculation used for this selection uses the DP distance, and uses a precise distance that takes into account temporal fluctuations.

In nonlinear matching, the DP (dynamic programming) method is a typical method. It is also called rubber matching because the matching operation is performed while expanding and contracting like rubber.

When calculating the distance between a standard pattern and an input pattern, DP matching partially matches the input side so that the distance between the two patterns is minimized, without making a one-to-one correspondence between the time series information of the two patterns. This is a method of matching while shifting.

Dialer 409 receives the voice number output from main selection processing section 407 and transmits a pulse corresponding to the voice number to the line.

Next, the operation of the system when the user issues "1" will be explained.

First, the user speaks into the microphone 401.

At the same time, the voice section detecting section 403 detects the beginning and end of "I" and stores the data output from the feature extracting section in this voice section.

Next, the normalization processing unit 404 normalizes the saved audio features on the time axis to create a feature pattern of 96 parameters.

Next, the preliminary selection processing unit 405 uses the configuration shown in FIG.
Calculate a preliminary selection distance based on the feature pattern of 0,
Ten feature patterns are selected from among them and recorded in a dictionary 408 of preliminary selection results.

Next, the 11 line selection processing unit 407 calculates the main selection distance based on the feature pattern created by the input voice and the 10 feature patterns included in the dictionary 408 of the preliminary selection results, using the configuration shown in FIG. The pattern number of the feature pattern that shows the minimum value is calculated, and the corresponding voice number (1
), the dialer 409 then outputs the received voice number (1
) sends a "1" signal to the line.

Here, the operation of the preliminary selection processing section 405 will be explained in more detail with reference to FIG.

The input pattern register 201 stores characteristic patterns created from input speech.

The dictionary memory 203 records registered IO voices and 30 patterns of characteristic bakuns.

The registered pattern number data memory 202 stores, for example, (11
,12,13,21,22,23,...,01,
02,031 is stored. In this case, the first digit is the voice number and the bottom one is the dark value number. When registering, for one voice Xl, Xl, X2 . X3(
7) This means that 3 patterns have been registered.

The preliminary selection simple distance calculation unit 204 targets all parameters of even channels, calculates the absolute value of the difference between the input feature pattern and the dictionary feature pattern, and calculates the preliminary selection simple distance by summing these values.

The preliminary selection distance of each feature pattern with respect to the input pattern is recorded in the distance data temporary stack memory 205 together with the pattern number.

Next, the minimum distance determination unit 206 detects the distance indicating the smallest value from this distance data, and outputs its pattern number.

The pattern number output at this time is not necessarily the pattern number belonging to voice number l+11.12.13), but it is regarded as a candidate and is stored in the preliminary selection result memory 207 as a preliminary selection result. Record.

At the same time, the output pattern number, for example 21,
is input to the matching section 210, and the matching section 210 reads out the voice number 2 corresponding to the pattern number 21 from the pattern number-voice number link data memory 2o9 and outputs it.

The voice number 2 output from the collation unit 210 is sent to the re-verification unit 21
1, and the re-verification unit 211 refers to the data stored in the pattern number-voice number link data memory 209 to find the pattern number (21) corresponding to the voice number 2.
, 22, 231 are output.

Next, the characteristic patterns corresponding to these pattern numbers +2L22.23) are deleted from the distance data temporary stack memory 205.

Then, as before, the pattern number obtained by the minimum distance determining section 206 is added to the preliminary selection result.

Such operations are repeated until the preliminary selection result becomes the lo pattern.

〔Effect of the invention〕

As explained above, according to the present invention, since pattern numbers corresponding to the same voice number are deleted, there is no need to select multiple patterns with the same voice content as selected feature patterns in preliminary selection. This greatly contributes to improving the recognition rate.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of the principle of the present invention. FIG. 2 is a block diagram showing a preliminary selection device according to an embodiment of the present invention. FIG. 3 is an example of the configuration of a main selection processing section after preliminary i-selection according to the present invention. FIG. 4 is a block diagram showing the system configuration of a speech recognition device which is the background of the present invention. FIG. 5 is a block diagram showing an example of a conventional preliminary selection device. In FIG. 2, 1... Preliminary selection means, 2... Distance calculation means, 3... Distance data temporary storage means, 4... Preliminary selection means 4. 5...Link data storage means, 201...Input pattern register, 202...Registered pattern number data memory, 203...
- Dictionary memory, 204... Simple distance calculation unit for preliminary selection, 205...
Distance data temporary stack memory, 206... Minimum distance determination section, 207... Preliminary selection result memory, 208... Counter, 209... Pattern number-voice number link data memory, 210... Verification section , 211... Re-verification section.

Claims (1)

[Claims] 1. Preliminary selection feature pattern storage means (1) for registering and storing in advance a group of a plurality of preliminary selection feature patterns corresponding to each input audio content; and an input made from the input audio content. distance calculation means (2) for calculating the distance between the feature pattern and each preliminary selection feature pattern stored in the preliminary selection feature pattern storage means (1); Distance data temporary storage means (3) for temporarily storing distance data; and preliminary selection means (4) for selectively outputting a preliminary selection feature pattern indicating the minimum distance among the distances stored in the distance data temporary storage means (3). ), and link data storage means (5) for storing in advance the relationship between each feature pattern and all feature patterns of the group to which the feature pattern belongs, and a link data storage means (5) for storing in advance the relationship between each feature pattern and all feature patterns of the group to which the selected and outputted feature pattern for preliminary selection belongs. output all feature patterns from the link data storage means (6) and input them to the distance data storage means (3), thereby
A preliminary selection device for speech recognition, characterized in that distance data corresponding to all feature patterns of a group to which the selectively outputted preliminary selection feature pattern belongs is deleted from the distance data storage means (3).