JPH0546196A - Speech recognition device - Google Patents

Abstract

PURPOSE:To correctly recognize an input speech to which a non-steady noise is added by learning plural noise patterns including non-steady noises from past input voice patterns. CONSTITUTION:A noise learning part 5 learns steady noises and non-steady noises from patterns judged as non-speech sections by a speech detection part 2 and a recognition part 3. Those are registered in a speech pattern storage part 6 and the registered noise patterns are used for recognition for a speech which is inputted next. Consequently, a high recognition ratio can be maintained even in environment where there are many non-steady noises.

Description

Detailed Description of the Invention

[0001]

TECHNICAL FIELD The present invention relates to a speech recognition apparatus, and more particularly to a noise learning method for a speech recognition apparatus used in an environment with a lot of non-stationary noise.

[0002]

2. Description of the Related Art It is known that learning a noise pattern is effective for a voice recognition device on the field, and has already been put to practical use.

A block diagram of the conventional speech recognition apparatus is shown in FIG. In the figure, first, the input signal a is converted into an input pattern b represented by a feature vector by the analysis unit 1. Next, the input pattern b is divided into a voice section and a non-voice section by the voice detection unit 2, the former is output to the recognition unit 3 as the input voice pattern c, and the latter is output to the noise learning unit 5 as the non-voice pattern d. ..

Next, the noise learning unit 5 extracts, for example, one section having the minimum power level from the non-voice pattern d,
Alternatively, the average of all sections is calculated and output to the recognition unit 3 as a noise pattern f.

Finally, the recognition unit 3 inputs the input voice pattern c.
And a pattern obtained by combining the noise pattern f before and after the standard pattern e, which is the output from the standard pattern storage unit 4, as a result, the category of the standard pattern having the highest similarity is output as the recognition result g. In addition, noise subtraction may be performed in which the noise pattern f is subtracted from the input voice pattern c during matching.

Conventionally, this kind of noise pattern is learned only from one pattern from a non-voice section as a result of voice detection on the assumption that it is stationary noise.

FIG. 6 (a) shows an example of a power waveform of an input pattern without non-stationary noise, and FIG. 6 (b) shows an example of a power waveform of an input pattern in which a tongue tap sound is added before vocalization and an exhalation sound is added after vocalization. (C) shows an example of the power waveform of the input pattern in which the ringing tone of the telephone is added from immediately before to immediately after the utterance.

In the conventional method, only the non-voice steady section SN is learned as a noise pattern in each of the power waveforms shown in FIGS. The cause of erroneous recognition in the actual field is that non-stationary noise is more common than stationary noise, and even though there are many applications used in environments where there is much non-stationary noise, noise learning in conventional speech recognition systems , Unsteady noise cannot be learned. Therefore, there is a problem that a high recognition rate cannot be maintained in the field.

[0009]

It is an object of the present invention to provide a speech recognition apparatus capable of learning non-stationary noise.

[0010]

First of all, although it is generally not expected that non-stationary noise will be repeated, in practice, similar non-stationary noise may be repeated while being used in a certain environment. Many.

For example, in an application of voice input from a telephone such as a banking service, voice input from the same speaker is performed several tens of times in the same environment from the time the telephone line is connected to the time the line is disconnected. During this period, for example, in an office where a ringing tone of a telephone or a running noise of a train is loud and frequently heard, these non-stationary noises are repeatedly mixed with the input voice.

Further, non-stationary noise peculiar to a speaker is repeatedly mixed, such as a person who has a habit of making a phone call by whispering a chair, or a person who has a loud tongue sound, an exhalation sound, or a nasal breath sound added before and after utterance. I often do it.

Next, when the same non-stationary noise is mixed, a certain person's specific word (voice that is strong against noise) is correctly recognized, but another specific word (voice that is weak against noise) is erroneously recognized. There are many. In other words, non-stationary noise added to speech that is strong against noise can be learned in reverse from the recognition result, and misrecognition can be prevented by using it when recognizing speech that is weak to non-stationary noise learned here. ..

[0014]

Therefore, according to the present invention, an analyzing section for converting an input signal into an input pattern, a voice detecting section for dividing the input pattern into an input voice pattern and a non-input voice pattern, and the non-voice pattern A noise learning unit for learning a noise pattern, a noise pattern storage unit for registering the noise pattern, a standard pattern storage unit for storing a standard pattern prepared in advance, the input voice pattern, the standard pattern, and the noise. A recognition unit that outputs a recognition result from the pattern, and the noise learning unit calculates a change amount of the feature vector from the non-voice pattern, and a calculation unit that calculates a moving average of the change amount for a predetermined frame. Then, the minimum section of the moving average is detected and used as a stationary noise pattern, and the section in which the moving average is larger than a preset value is detected. Speech recognition device is obtained which comprises a means for the non-stationary noise pattern this by.

[0015]

Embodiments of the present invention will be described below in detail with reference to the drawings.

FIG. 1 is a block diagram of an embodiment of the present invention, and the same parts as in FIG. 2 are designated by the same reference numerals. Figure 3
5 is a processing flowchart showing the operation of each unit of FIG. 1, and the configuration and operation of FIG. 1 will be described with reference to FIGS.

First, the input signal a is converted by the analysis unit 1 into an input pattern b represented by a feature vector. The input pattern b is divided into a voice section and a non-voice section by the voice detection unit 2 (step 21), the former is output to the recognition unit 3 as the input voice pattern c (step 22), and the latter is noise-learned as the non-voice pattern d. Output to the unit 5 (step 23). Here, the voice detection parameter is set so that the input voice pattern sufficiently includes the true voice section,
A hangover section may be added.

Next, in the noise learning section 5, the non-voice pattern d
The average of the sections in which the change of the feature vector is small is calculated (steps 51 and 52) and is output to the noise pattern storage unit 6 as the stationary noise pattern h (steps 53 and 5)
4). If there is a section in which the change of the feature vector is large in the non-voice pattern d, the section is output to the noise pattern storage unit 6 as the non-stationary noise pattern i (steps 55 and 54).

In FIG. 6, it is assumed that the voice is accurately detected and the non-voice pattern is present except in the "VOICE" section, and that the variation amount of the feature vector is substantially the same as the power waveform. In this case, the "SN" section in which the change in the feature vector is small is learned as the stationary noise pattern h, and the section (N1, N2, N3, N3 ') in which the change is large is learned as the non-stationary noise pattern i.

Next, when some noise patterns have already been registered in the noise pattern storage unit 6, the noise pattern (N1,
N2, N3, N3 ') with a high degree of similarity (N8, N9
If there is any), these (N8, N9) are discarded (steps 61, 62).

Since N3 and N3 'have a high degree of similarity,
N3 'is not registered.

Furthermore, when the number of noise patterns registered in the noise pattern storage unit 6 exceeds a predetermined number (M) (step 63), the noise pattern registered the earliest (N7). ) Is discarded (step 6)
4).

Next, in the recognition section 3, the input voice pattern c and the standard pattern e which is the output from the standard pattern storage section 4 are inputted.
And matching with the noise pattern f which is the output from the noise pattern storage unit 6.

As the matching method, for example, first, the matching with the noise pattern is started from the input voice pattern.
The noise pattern includes a combination of a stationary noise pattern and a non-stationary noise pattern.After the noise pattern with the highest similarity, matching is performed between the standard pattern and the subsequent section of the input speech pattern, and finally It also matches the noise pattern with the section from the continuation to the end of the input speech pattern.

The category of the standard pattern having the highest similarity as a result of matching is output as the recognition result g. In addition, when performing matching with a standard pattern, noise subtraction may be performed in which a stationary noise pattern is subtracted from an input voice pattern.

Next, when the recognition result is correct (step 31), a matching back trace with the correct standard pattern is performed (step 32), and the section of the input voice pattern is checked in correspondence with the standard pattern. Of the input voice pattern corresponding to the noise pattern is determined to be a non-voice segment, and this is output to the noise learning unit 5 as a non-voice pattern j (step 33).

In FIG. 6, if voice detection is not performed accurately and the "DETECT" section becomes the input voice pattern, this section becomes the target of matching, and the matching back in the case of the input voice pattern of FIG. 6B is performed. The trace is shown in FIG. Here, the section corresponding to the standard pattern is t2 to t3, and this section is determined to be a true voice section. On the contrary, the sections corresponding to the noise pattern are t1 to t2 and t3 to t4, and this section is determined to be a non-voice section and is output to the noise learning unit as a non-voice pattern j.

Next, in the noise learning section 5, the non-voice pattern j
If there is a small change in the feature vector, the average is taken and the noise pattern storage unit 6 is used as the stationary noise pattern h.
(Steps 51 to 54). Also, if there is a section in which the change of the feature vector is large in the non-voice pattern j, the section is output to the noise pattern storage unit 6 as the non-stationary noise pattern i (steps 51, 52, 55).

In FIG. 6, assuming that the amount of change in the feature vector is almost the same as the power, there is no interval where the change is small, so there is no stationary noise pattern, and there is a non-stationary noise pattern (N1, N2). N3, N3 ')
Will be learning.

Finally, the method of discarding and registering the noise pattern registered in the noise pattern storage unit 6 is the same as described above.

As described above, a noise pattern including non-stationary noise is learned from past input patterns used in the same environment (place, person, etc.), and matching is performed using these noise patterns at the next recognition. It can be performed.

[0032]

As described above, according to the present invention, non-stationary noise can be learned and recognized in addition to stationary noise. Therefore, a plurality of noise patterns including non-stationary noise are used. There is an effect that matching can be performed, and voice recognition can be performed with a high recognition rate in an application having a lot of non-stationary noise.

[Brief description of drawings]

FIG. 1 is a block diagram of an embodiment of the present invention.

FIG. 2 is a block diagram of a conventional voice recognition device.

FIG. 3 is a processing flowchart of a voice detection unit 2 and a noise learning unit 5.

FIG. 4 is a flowchart showing a storage procedure of a noise pattern storage unit 6.

FIG. 5 is a processing flowchart of the recognition unit 3.

6A is a power waveform diagram of an input pattern having no non-stationary noise, and FIGS. 6B and 6C are power waveform diagrams of an input pattern in which non-stationary noise is mixed.

FIG. 7 is a diagram showing a matching back trace of the input voice pattern of FIG. 6 (b) with a standard pattern and a noise pattern.

[Explanation of symbols]

 1 analysis unit 2 voice detection unit 3 recognition unit 4 standard pattern storage unit 5 noise learning unit 6 noise pattern storage unit

Claims (2)

[Claims] 1. An analysis unit for converting an input signal into an input pattern, a voice detection unit for dividing the input pattern into an input voice pattern and a non-input voice pattern, and a noise learning unit for learning a noise pattern from the non-voice pattern. And a noise pattern storage unit for registering the noise pattern, a standard pattern storage unit for storing a standard pattern prepared in advance, the input voice pattern, the standard pattern, and the noise pattern. Including a recognition unit,
The noise learning unit calculates a change amount of the feature vector from the non-voice pattern, a unit that calculates a moving average of a predetermined frame of the change amount, and detects a minimum section of the moving average and detects the minimum interval. A speech recognition apparatus comprising: a stationary noise pattern, and means for detecting a section in which the moving average is larger than a preset value and setting this as a non-stationary noise pattern. 2. The noise learning unit further comprises means for newly learning a noise pattern from a section of an input voice pattern corresponding to the noise pattern as a result of matching processing which is recognition processing performed by the recognition unit. The voice recognition device according to claim 1, comprising: