JPH071437B2 - Voice recognizer - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Industrial field of application] The present invention converts uttered speech into a standard pattern and DP (Dynamic).
TECHNICAL FIELD The present invention relates to a speech recognition apparatus that performs recognition by determining a standard pattern that gives a minimum degree of difference, and particularly to a speech recognition apparatus that is not easily affected by environmental noise.

[Conventional technology]

In a conventional voice recognition device, a uttered voice is input from a microphone, the amplitude (including power) and spectrum of a voice signal wave converted into an electric signal are tested to detect the voice, and the voice of the section is detected. I am aware of

And, usually, the point where the amplitude level exceeds a certain threshold value and the point where it falls below it are used as the start and end points, or the point where the spectrum changes rapidly in the vicinity of the above point is set as the start and end points.
The recognition process is performed for the voice section. The number "1" (/ i
A pause interval (silence) is observed between words such as cni /) and Sapporo (/ sapporo /), and between continuously uttered words. Note that there is a method of including a pause section in a word and a method of not including it in the voice section.

On the other hand, the place to input voice data is not only in a quiet office but also in many places such as in a factory or outdoors where noise is generated from various machines. In addition, a speech recognition apparatus generally uses a noise canceling close-talking microphone to improve noise immunity, but this is not enough. If the noise level exceeds the threshold for voice detection, or if the spectrum of the noise itself changes before and after the beginning and end of true utterance, a situation occurs in which voice detection is erroneous. Also, if noise is superimposed on a pause interval between words or between words and the voice detection section is mistaken, the pattern length becomes apparently long and it becomes difficult to match with the standard pattern, and the recognition result of the entire utterance is erroneous. become. Then, when the noise is not so high, the influence of the noise can be reduced to some extent by raising the threshold and removing the pause section in the word from the voice section. However, when the noise amplitude or spectrum changes significantly in a short time, that is, when it is non-stationary noise, the threshold is set higher than the peak value of the noise, and this time the start and end of the utterance and Since it becomes difficult to detect a small amplitude portion or a consonant portion near the pause point, the recognition performance is significantly reduced, and this method is not practical.

In order to reduce the effect of this voice detection error,
The end is not fixed and the beginning and end have a width.
There is a so-called start / end free recognition method. Then, the recognition of the start / end free is performed by comparing and collating the voice patterns of the sections of all the combinations of the start and end candidate points, and determining the most likely recognition result as the final result. Will be realized. One example thereof is described in detail in Japanese Patent Application No. 61-31179. Although it is possible to reduce the detection error of the beginning and end of the voice section by the end-free recognition, there is no problem about the noise included in the pause section of the utterance and the noise included in the voice section. It is often the case that correct recognition results cannot be obtained due to ineffectiveness.

[Problems to be Solved by the Invention]

In the above-mentioned conventional voice recognition method, when voice detection is performed based on the amplitude level or spectrum change of the voice, and voice recognition is performed at the start / end free, an error due to addition when voice is mixed in the pause section during utterance There was a problem that it was not solved yet.

[Means for Solving the Problems]

The speech recognition apparatus according to the present invention is a speech recognition apparatus which performs DP matching of uttered speech with a standard pattern and obtains a standard pattern that gives a minimum degree of difference. DP to calculate
When the DP matching path has a matching part and passes through the rest position on the standard pattern side, finds a temporary rest point that gives the minimum value of the difference near the position on the input pattern side corresponding to it, and determines the difference As a condition, there is provided means for continuing DP matching between the portion after the rest position of the standard pattern and the portion in the vicinity of the point preceding the temporary rest point of the input pattern without end points.

[Action]

In the present invention, since the end-point-free matching is performed even before and after the position of the input pattern corresponding to the rest position on the standard pattern side, it is not necessary to accurately detect the rest period of the input pattern.

〔Example〕

 Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

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

In the figure, 1 is an input section for inputting an audio signal wave s, 2 is a standard pattern memory for receiving the output of the input section 1, 3
Is an input pattern memory whose input is the output of the input unit 1, and 4 is an input pattern 1 from the input pattern memory 3 and a standard pattern j from the standard pattern memory 2 is input, and a difference degree proportional to the standard pattern time length is shown. In the DP matching section for calculation, these are temporary rest points that give the minimum value of the difference in the vicinity of the corresponding position on the input pattern side when the DP matching path passes through the rest position on the standard pattern side. A means for continuing DP matching between the portion after the rest position of the standard pattern and the portion in the vicinity of the point preceding the temporary rest point of the input pattern using the degree of difference as a boundary condition is constituted.

The operation of the embodiment shown in FIG. 1 will be described below.

First, the input unit 1 detects a section in which the amplitude level of the input voice signal wave s is higher than a predetermined threshold as a voice section and converts it into a time-series pattern of characteristic parameters. If there is a pause section in the word, its position is also detected. Then, at the time of registration, the time-series pattern and the rest position are stored in the standard pattern memory 2. At the time of recognition, the time series patterns are temporarily stored in the input pattern memory 3.

Next, the DP matching unit 4 calculates a difference degree proportional to the standard pattern time length based on the input pattern i output from the input pattern memory 3 and the standard pattern j output from the standard pattern memory 2. And the standard pattern memory 2
Also, the rest position information q of the standard pattern is displayed by the DP matching unit 4
Be instructed to.

FIG. 2 is a diagram for explaining the process of DP matching used in the explanation of the operation in FIG. 1, and is an explanatory diagram in which the horizontal axis represents the input pattern i and the vertical axis represents the standard pattern j.

When the DP matching path D passes through the rest position Q on the standard pattern side, a point P ₁ (temporary rest point) that gives the minimum difference in the corresponding neighborhood K ₁ on the input pattern side is obtained. This point P ₁
As a boundary condition, the DP matching is continued between the partial pattern after the rest position Q on the standard pattern side and the neighborhood K _{2 of} the point preceding the temporary rest point of the input pattern, free of end points. In FIG. 2, it is shown that the DP matching continued from the point P ₂ as a result of the end point-free matching. point
This means that the interval between P ₁ and P ₂ is treated as the pause interval on the input pattern side.

Then, even if noise is mixed in the pause interval and the voice detection is erroneous, the noise component of the input pattern is skipped and DP matching is performed. Here, although the rest position and the number thereof are different for each standard pattern, by repeating the above calculation,
The difference between the start end and the end of the standard pattern and the input pattern is obtained, and finally the standard pattern with the smallest difference normalized by the standard pattern length J is output as the standard result R.

〔The invention's effect〕

As described above, according to the present invention, since the end-point-free matching is performed even before and after the position of the input pattern corresponding to the rest position on the standard pattern side, it is not necessary to accurately detect the rest period of the input pattern, and the utterance Since it is possible to solve the problem that noise is superimposed on the pause section and the input pattern becomes long to cause erroneous recognition, there is an effect that the recognition performance in normal time can be maintained even when there is unsteady noise. Further, the present invention can be applied not only to the pause section in a word but also to the pause section between words, so that it is also effective in continuous word recognition.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention, and FIG. 2 is an explanatory diagram for explaining the process of DP matching used for the operation explanation of FIG. 1 ... Input section, 2 ... Standard pattern memory, 3 ... Input pattern memory, 4 ... DP matching section.

Claims (1)

[Claims] 1. A speech recognition apparatus for recognizing a uttered voice by performing DP matching with a standard pattern and obtaining a standard pattern giving a minimum degree of difference, and determining a degree of difference proportional to a standard pattern time length. It has a DP matching part for calculation, and when the DP matching path passes through the rest position on the standard pattern side, finds a temporary rest point that gives the minimum value of the difference near the position on the input pattern side corresponding to this With a degree as a boundary condition, a means is provided for continuing the DP matching without end points between the portion after the rest position of the standard pattern and the portion in the vicinity of the point preceding the temporary rest point of the input pattern. And a voice recognition device.