JPH0651792A - Speech recognizing device - Google Patents

Abstract

PURPOSE:To correctly recognize a speech even if a standard pattern is similar by providing a threshold value for the distance between corresponding frames of an input pattern and the standard pattern. CONSTITUTION:This speech recognizing device consists of a microphone 21 which inputs the speech, an amplifier 22, a speech analytic part 23 which outputs speech parameters on the basis of the speech, a feature pattern extraction part 24 which extracts the input pattern divided into frames from the speech parameter, a standard pattern 26, an inter-frame distance calculation part 25 which calculates the distance between the corresponding frames of the input and standard patterns, a threshold value holding part 272 which holds the threshold value of the distance between the frames, a weighting part 273, an inter- pattern distance holding part 274 which adds distances between weighted and unweighted frames to all the frames and holds the result as the inter-pattern distance between the input and standard pattern, an inter-pattern distance comparison part 28 which detects the most similar pattern to the input pattern according to the inter-pattern distance, and a minimum pattern distance holding part 29.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device for controlling various devices by voice.

[0002]

2. Description of the Related Art In recent years, a voice recognition device for recognizing a voice by performing pattern matching between an input pattern representing a feature of an input voice and a standard pattern representing a feature of a standard voice has been actively developed. This type of device has been put to practical use.

FIG. 1 shows the configuration of a conventional voice recognition device.
The operation will be described based on the configuration.

First, the voice input from the microphone 11 is amplified by the amplifier 12 to the extent that the amplitude is not saturated,
The voice analysis unit 13 outputs voice parameters. The characteristic pattern extraction unit 14 determines 10 msec from the voice parameter.
Extract an input pattern divided into frames at intervals of ˜300 msec. The inter-frame distance calculation unit 15 calculates a distance (similarity calculation) between corresponding frames of the input pattern and the value of the standard pattern 16, and divides the output of the distance calculation into all frames (1 in the case of word recognition). (For words)
Is added to and is held in the inter-pattern distance holding unit 17 as the inter-pattern distance between the input pattern and the standard pattern. The value of the inter-pattern distance holding unit 17 is compared with the minimum inter-pattern distance obtained from the inter-pattern distance holding unit 17 in the inter-pattern distance comparing unit 18,
Information indicating the smaller inter-pattern distance and the type of voice associated with the standard pattern at that time is rewritten and held in the minimum inter-pattern distance holding unit 19. After repeating the same processing for all the standard patterns, the voice information held in the minimum inter-pattern distance holding unit 19 is output as a recognition result signal.

However, in such a method in which the distance between patterns is simply defined as the sum of the distances between the corresponding frames of the input pattern and the standard pattern, and the same comparison is performed with all the standard patterns, the standard pattern is used. If there are few frames that are similar to each other and show differences in characteristics, the difference in the overall pattern distance between the input pattern and the standard pattern becomes relatively small, making it difficult to distinguish between these standard patterns. was there.

[0006]

SUMMARY OF THE INVENTION The present invention has been made in view of the above problems, and provides a voice recognition device capable of avoiding erroneous recognition even if there are a plurality of similar standard patterns. is there.

[0007]

A voice recognition device of the present invention is a device for recognizing a voice by performing pattern matching between an input pattern representing a feature of an input voice and a standard pattern representing a feature of a standard voice. Therefore, a voice analysis unit that analyzes an input voice and outputs a voice parameter, a feature pattern extraction unit that extracts an input pattern divided into frames from the voice parameter, and a standard that stores a plurality of standard patterns divided into frames A pattern memory, an inter-frame distance calculation unit that calculates a distance between corresponding frames of the input pattern and the standard pattern, a threshold value holding unit that holds a threshold value of the distance between the frames, and the inter-frame distance calculation unit Weights the output of the inter-frame distance comparison unit that compares the output of A weight applying section, and the pattern distance holder, based on the pattern distance between the pattern distance maintaining unit consists of a comparison means for detecting the most similar reference pattern in the input pattern, the pattern distance holding portion,
When the output of the inter-frame distance calculation unit exceeds the value of the threshold value holding unit, the weighting unit gives a weight to the output of the inter-frame distance calculation unit and does not give the weighted output or the weighted output. The output is added to all the divided frames and held as the inter-pattern distance between the input pattern and the standard pattern.

[0008]

The speech recognition apparatus of the present invention detects a frame position where a feature difference occurs between these similar standard patterns even when there are a plurality of similar standard patterns, and the input pattern of this frame position and the standard pattern are detected. Since the threshold of the distance between patterns is set low and weighting is performed when this threshold is exceeded, the difference in the distance between the above frames is emphasized and a significant difference is found in the distance between the input pattern and the standard pattern. You can get out.

[0009]

FIG. 2 shows the configuration of an embodiment of a voice recognition device according to the present invention.

In FIG. 2, a microphone 21 is for inputting a voice to the voice recognition device, and an amplifier 22 is for amplifying the voice from the microphone 21 to the extent that the amplitude is not saturated. The voice analysis unit 23 outputs a voice parameter from the amplified voice. The characteristic pattern extraction unit 24 determines 10 msec to 300 m from the voice parameter.
An input pattern divided into frames at intervals of sec is extracted. Here, the reason why the frame interval is set to 10 msec or more is that there is no meaning because the input pattern does not change even if the frame is divided into too small parts. Also, 3
The reason why it is set to 00 msec or less is that it is difficult to capture the features because the division is too coarse. The inter-frame distance calculation unit 25 uses the input pattern and the standard pattern 26.
Calculate the distance between the corresponding frames with the value of.

The inter-pattern distance calculating means 27 is an essential part of the present invention, and includes an inter-frame distance comparing section 271, a threshold value holding section 272, a weighting section 273, and an inter-pattern distance holding section 27.
It consists of 4 sections and is explained below.

The threshold value holding unit 272 holds the threshold value of the distance between the frames corresponding to the input pattern and the value of the standard pattern 26. The weighting unit 273 weights the output of the inter-frame distance calculation unit 25. The inter-frame distance comparison unit 271 uses the inter-frame distance calculation unit 25.
Is compared with the value of the threshold holding unit 272.
The inter-pattern distance holding unit 274 compares the threshold value preset in the threshold value holding unit 272 with the output of the inter-frame distance calculation unit 25, and when the output exceeds the threshold value, the weighting unit 273 performs the above-mentioned operation. Weight the output,
If the output does not exceed the threshold value, the weighted output or the unweighted output is added to all the divided frames (for example, one word) to obtain the input pattern. This is held as the inter-pattern distance of the standard pattern.

The minimum inter-pattern distance holding unit 29 holds information indicating the type of voice associated with the standard pattern having the minimum inter-pattern distance. The inter-pattern distance comparison unit 28 compares the output of the inter-pattern distance calculation unit 27 and the value of the minimum inter-pattern distance holding unit 29, and the value with the smaller inter-pattern distance is the minimum inter-pattern distance holding unit 29.
To hold.

The operation of the inter-pattern distance calculation means 27, which is the main part of the present invention, will be described below with reference to the block diagram of the embodiment of the present invention in FIG. 2 and the flow chart in FIG.

Here, it is assumed that the following settings are made as a premise for the processing of the inter-pattern distance calculation means 27.

First, the standard pattern 26 is the prefecture name 4
Seven items (“Hokkaido”, ..., “Tokyo”, “Osaka”, “Okayama”, “Wakayama” ...) are registered. Also, the word voice "Okayama" is input from the microphone. The frame interval is 125 ms.
ec, and all 47 standard patterns consist of 8 frames. The threshold value Th (h is 0 to 7) has similar standard patterns of "Okayama" and "Wakayama". Therefore, the first two frames are ⅕ of the normal one, that is, T0 = 20, T.
1 = 20, T2 = 100, T3 = 100, T4 = 10
It is set to 0, T5 = 100, T6 = 100, T7 = 100. Further, the weight adding unit 273 weights the output of the inter-frame distance calculating unit 25 twice.

First, in S1, the preset Th corresponding to the inter-frame distance value output from the inter-frame distance calculating section 25 is stored in the threshold holding section 272. Next, in S2, the distance between one frame of the input pattern and the standard pattern is calculated. Here, the distance value is set to Df. In S3, the size of the distance value Df and the threshold value Th are compared, and in the case of Df> Th, the distance value is weighted in S4. Since the weighting factor α = 2, the new inter-frame distance value is updated to Df = Df × 2 and then added to the inter-pattern distance holding unit 274 in S5.
If the inter-pattern distance value of the current frame is expressed as Dp (i),
Dp (i) = Dp (i-1) + Df. If Df <Th as a result of the size comparison in S3, Df is added as it is. The procedure from S2 onward is repeated until the distance calculation for all frames is completed, and finally the pattern distance between the input pattern and the standard pattern is determined.

In comparison with the standard pattern "Okayama",
The output of the inter-frame distance calculation unit is F0 (frame 0) =
15, F1 = 15, F2 = 20, F3 = 15, F4 = 2
0, F5 = 20, F6 = 15, F7 = 20, 1
5 + 15 + 20 + 15 + 20 + 20 + 15 + 20 = 14
0, and the distance between patterns is 140.

In comparison with the standard pattern "Wakayama",
The output of the inter-frame distance calculation unit is F0 (frame 0) =
30, F1 = 25, F2 = 15, F3 = 10, F4 = 2
If 0, F5 = 10, F6 = 15, F7 = 10, then F
Since 0 (frame 0) and F1 exceed the threshold, the calculation is 30 × 2 + 25 × 2 + 15 + 10 + 20 + 10 + 15
+ 10 = 190, and the distance between patterns is 190.

In this way, the standard pattern is "Okayama".
Even if the pattern is similar to "Wakayama", it is clearly distinguished and recognized as "Okayama".

In the above example, if no weight is given,
The distance between patterns of "Okayama" remains 125,
The pattern-to-pattern distance of "Wakayama" becomes 135, the pattern-to-pattern distance of "Wakayama" becomes small, and it is erroneously recognized as "Wakayama".

According to the apparatus of the present invention described above, the difference between the corresponding frames of the input pattern and the standard pattern is emphasized by setting the threshold value for determining the weighting process to a low level, and the pattern of the input pattern and the standard pattern is emphasized. Since a significant difference can be produced in the inter-distance, even if a plurality of standard patterns are similar and there are few frames in which there is a difference, the inter-pattern distance difference can be increased and the discrimination ability can be improved.

[0023]

According to the speech recognition apparatus of the present invention, even if there are a plurality of similar standard patterns, the frame position where a difference in characteristics occurs between these similar standard patterns is detected, and the input pattern of this frame position is detected. Since the threshold of the distance between the standard pattern and the standard pattern is set low and weighting is performed when the threshold is exceeded, the difference between the distances between the frames is emphasized, and the pattern distance between the input pattern and the standard pattern is significant. You can make a difference. Therefore, erroneous recognition of similar words can be reduced and the recognition performance can be improved.

[Brief description of drawings]

FIG. 1 is a configuration diagram of a conventional voice recognition device.

FIG. 2 is a configuration diagram of an embodiment of a voice recognition device of the present invention.

FIG. 3 is a flowchart showing a control procedure of an inter-pattern distance calculation means which is a main part of the voice recognition device of the present invention.

[Explanation of symbols]

 11 Microphone 12 Amplifier 13 Speech Analysis Unit 14 Feature Pattern Extraction Unit 15 Inter-frame Distance Calculation Unit 16 Standard Pattern 17 Inter-pattern Distance Holding Unit 18 Inter-pattern Distance Comparison Unit 19 Minimum Inter-pattern Distance Holding Unit 21 Microphone 22 Amplifier 23 Speech Analysis Unit 24 Feature pattern extraction unit 25 Inter-frame distance calculation unit 26 Standard pattern 27 Inter-pattern distance calculation means 271 Inter-frame distance comparison unit 272 Threshold holding unit 273 Weighting unit 274 Inter-pattern distance holding unit 28 Inter-pattern distance comparison unit 29 Minimum inter-pattern distance Holding part

Claims (1)

[Claims] 1. A voice recognition device for recognizing a voice by performing pattern matching between an input pattern representing a feature of an input voice and a standard pattern representing a feature of a standard voice, and analyzing the input voice and outputting a voice parameter. A voice analysis unit, a feature pattern extraction unit for extracting an input pattern divided into frames from the voice parameter, a standard pattern memory storing a plurality of standard patterns divided into frames, the input pattern and the standard pattern The inter-frame distance calculation unit that calculates the distance between the corresponding frames, the threshold holding unit that holds the threshold value of the distance between the frames, and the output of the inter-frame distance calculation unit and the value of the threshold holding unit are compared. An inter-frame distance comparison unit, a weighting unit for weighting the output of the inter-frame distance calculation unit, and an inter-pattern distance The inter-pattern distance holding unit outputs the output of the inter-frame distance calculation unit. The inter-pattern distance holding unit includes a holding unit and a comparing unit that detects a standard pattern most similar to the input pattern based on the inter-pattern distance holding unit. When the value exceeds the value of the threshold value holding unit, the weighting unit gives a weight to the output of the inter-frame distance calculation unit, and the output with the weight or the output without the weight is applied to all the divided frames. A voice recognition device, characterized in that it is added to each other and held as an inter-pattern distance between the input pattern and the standard pattern.