JP2856429B2 - Voice recognition method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Object of the Invention] (Industrial application field) The present invention relates to a speech recognition method used in a speech recognition device, and particularly to a technique for rejecting a speech that is not to be recognized when it is input. The present invention relates to an improved speech recognition system.

(Prior art) The input and output of information by voice is highly natural for humans and is excellent as a man-machine interface.
Conventionally, various studies have been made.

As a speech recognition method for this purpose, a standard pattern created by performing smoothing or differentiation processing on a learning speech pattern collected in advance, and an input speech obtained by analyzing the input speech There is one that recognizes an input voice by calculating a similarity between a pattern and a smooth pattern (see Japanese Patent Application No. 62-252108).

In such speech recognition, whether or not the inputted speech is the speech to be recognized has been determined using the similarity calculated between the standard pattern and the smoothed pattern of the input speech pattern. .

(Problems to be Solved by the Invention) As described above, the determination as to whether or not the voice is the recognition target voice is performed using the similarity calculated between the standard pattern and the smooth pattern of the input voice pattern. , The similarity will be higher even if the non-recognized speech is input,
There was a problem of misrecognition.

Here, this will be described in detail with reference to FIG. For example, if the learning voice pattern is a pattern as shown in FIG. 3A and the smooth pattern of the input voice pattern not to be recognized is a pattern as shown in FIG. The calculated similarity is as shown in FIG. 3 (c), and a high level of similarity is obtained. For this reason, erroneous recognition occurs, and it is impossible to reject (reject) even if a voice not to be recognized is input.

Accordingly, it is an object of the present invention to provide a speech recognition method that enables rejection of a non-recognized speech with high accuracy without erroneous recognition.

[Configuration of the Invention] (Means for Solving the Problems) The present invention provides an input voice pattern obtained by analyzing input voice and a standard pattern created based on a learning voice pattern collected in advance. By calculating the degree of similarity or difference between and, in a voice recognition system that recognizes the input voice, using a plurality of filters that perform at least smoothing processing and differentiation processing on the learning voice pattern, Create a standard pattern,
Calculating a similarity or a difference between the axis of the standard pattern obtained by the differential processing and the differential pattern obtained by differentiating the input voice pattern, and determining whether or not the voice is a recognition target voice. And

(Action) By differentiating both the learning voice pattern and the input voice pattern to further emphasize the features of the voice pattern,
Since the magnitude of the similarity is sensitive to the difference in the voice pattern, when a voice not to be recognized is input, it can be rejected with high accuracy without erroneous recognition.

Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

FIG. 1 is a schematic configuration diagram of a speech recognition device configured by applying a speech recognition system according to the present invention. That is, a voice input from a voice input unit (not shown) is converted into an electric signal and fetched, an acoustic analysis is performed by an acoustic analysis unit 1 including a band-pass filter, and a word voice section is detected by a voice section detection unit 2. Is detected. From the input voice pattern detected in the voice section, the sample point extraction unit 3 extracts a predetermined number of sample points obtained by equally dividing the voice section in the time direction, and obtains a sample represented by (the number of feature vectors × the number of sample points). Find a pattern. The sample patterns obtained in this manner are collected by a predetermined number for each category to be recognized and stored in the voice pattern storage unit 4.

The standard pattern 6 is created by a plurality of filters for executing at least a smoothing process and a differentiation process on the sample pattern stored in the voice pattern storage unit 4, for example, an orthogonalization time filter including a plurality of orthogonalization time filters. Performed by part 5.

Here, the learning voice patterns collected in the voice pattern storage unit 4 include, for example, j (= 1, 2, to 16)
Is a data sequence collected over 18 sample points, which is composed of 16 acoustically analyzed feature vectors represented by the following, and divides the voice section into 17 equal parts as k (= 0, 1, 2, ~ 17). It will be described as given.

Now, the orthogonalization time filter unit 5 calculates the m-th learning voice pattern collected for each category i by three a
_{When m} (j, k) is set, the standard pattern 6 is created as follows.

(1) First, a learning voice pattern a _m (j,
k), the average pattern A (j, k) [I = 1,2, 〜16, k = 0,1,2, 〜17].

(2) Thereafter, using the average pattern A (j, k) obtained as described above, b1 (j, k) = A (j, k-1) + 2 * A (j, k) + A (j, k + 1) ... (2) [1 = 1,2, ~ 16, k = 1,2, ~ 16] The first axis b1 (j, k) of the standard pattern is obtained by the following calculation.
This is registered in the standard pattern 6. This standard pattern b1
(J, k) is obtained as a result of smoothing the average pattern A (j, k) in the time axis direction, and is registered as data of the first axis which is a reference of the standard pattern 6.

(3) Thereafter, using the average pattern A (j, k), b2 (j, k) = − A (j, k−1) hA (j, k + 1) (3) [j = 1,2 , ~ 16, k = 1,2, ~ 16] to obtain the second axis b2 (j, k) of the standard pattern,
After normalizing this, it is registered in the standard pattern 6. The standard pattern b2 (j, k) is obtained by differentiating the average pattern A (j, k) in the time axis direction.

The standard pattern 6 is created by repeatedly executing the above processes (1) to (3) for each category.

Even if the processing procedure by the orthogonalization time filter unit 5 is changed as follows, a substantially equivalent standard pattern 6 can be created. That is, (1) From the collected learning voice patterns a _m (j, k) The first axis b1 (j, k) of the standard pattern is obtained as [j = 1,2, 〜16, k = 1,2, 〜16] and registered in the standard pattern 6.

(2) Then, from the learning voice pattern a _m (j, k) The second axis b2 (j, k) of the standard pattern is determined as [j = 1,2, 〜16, k = 1,2, 〜16] and registered in the standard pattern 6.

Such processes (1) and (2) are repeatedly executed by the number of categories. That is, as described above, without first calculating the average pattern A (j, k), the first standard pattern of the standard pattern smoothed in the time axis direction from the collected predetermined learning voice pattern a _m (j, k). The axis b1 (j, k) and the second axis b2 (j, k) of the standard pattern differentiated in the time axis direction may be directly calculated.

By the way, in the above description, an example in which up to two axes are obtained as the standard pattern 6 has been described. However, three or more axes of the standard pattern may be created by performing secondary differentiation. In this case, it is sufficient to use, for example, 20 or more sample points extracted as the learning voice patterns instead of the 18 points described above. In this case,
For example, b1 (j, k) = A (j, k-2) + 4 * A (j, k-1) + 6 * A (j, k) + 4 * A (j, k + 1) + A (j, k + 2) ... ( 6) [j = 1,2, j16, k = 1,2, 〜16] to obtain the first axis b1 (j, k) of the standard pattern, and b2 (j, k) = − A (j, k−2) −2 * A (j, k−1) + 2 * A (j, k + 1) + A (j, k + 2) (7) [j = 1,2, 〜16, k = 1,2, 〜 16], the second axis b2 (j, k) of the standard pattern may be obtained. And the third of the second derivative standard pattern
For the axis b3 (j, k), b3 (j, k) =-A (j, k-2) -2 * A (j, k-1) + 3 * A (j, k) -2 * A (j , k + 1) -A (j, k + 2) (8) [j = 1,2, 〜16, k = 1,2, 〜16].

Next, similarity calculation at the time of speech recognition will be described. The similarity calculating unit for recognition 7 calculates all the axes of the standard pattern 6 created as described above and the sample pattern W of the input voice V in the smoothing pattern creating unit 8 as X (j, k) = W (j, k). -1) + 2 * W (j, k) + W (j, k + 1) (9) Smoothed pattern smoothed as [j = 1,2, -16, k = 1,2, -16] Between X Calculates the similarity between the standard pattern b _{i, r} of the category i _, and determines the judgment unit 9 according to the similarity.
To recognize the input voice V.

The standard pattern b _{i, r} category i has been pre-normalized, K _i denotes the number of standard patterns of category i (number of axes). Also, (•) indicates the inner product and {} indicates the norm.

Next, a method for determining whether or not an input voice is a recognition target word will be described. For example, a case will be described in which the recognition result for the input voice V is category I in the recognition processing. Similarity calculation unit for rejection 10
Is obtained by using the second axis of the standard pattern 6 of category I created as described above and the sample pattern W of the input voice V in the differential pattern creating unit 11 as Y (j, k) =-W (j, k-1). + W (j, k + 1) (11) between the differential pattern Y subjected to the differential processing as [j = 1,2, -16, k = 1,2, -16] Calculates the similarity between the standard pattern 6 of the category I and the second axis b _{I, 2,} and determines whether or not the input voice V is a recognition target word in the determination unit 9 according to the similarity value. Is determined. When it is determined that the input voice V is a word not to be recognized in the determination based on the similarity value, the category I of the recognition result obtained in the above-described recognition processing is rejected, and a request for re-speaking is performed as necessary. Is performed.

In this way, according to the present method, the rejection process is performed based on the similarity obtained between the differential pattern of the sample pattern of the input voice and the second axis of the standard pattern 6 of the category of the recognition result obtained by the recognition process. By differentiating the input voice, the similarity value for rejecting responds sensitively to the difference between the input voices, and the input voice can be used as it is for the similarity calculation, or can be smoothed before the similarity calculation. Words not to be recognized can be rejected with higher accuracy than rejection processing based on the rejection similarity calculation value obtained using the rejection similarity calculation value, and the practical effect is great.

Here, this will be described in detail with reference to FIG. As in the above-described example, for example, it is assumed that the learning voice pattern is a pattern as shown in FIG. 3 (a) and the smooth pattern of the input voice pattern not to be recognized is a pattern as shown in FIG. 3 (b). Then, the differential patterns of these two patterns are respectively shown in FIGS. 3 (d) and (e), and the similarity calculated between these two differential patterns is as shown in FIG. 3 (f), and the similarity of almost zero is obtained. Becomes Therefore, there is no erroneous recognition as in the related art.

In this way, by differentiating both the learning voice pattern and the input voice pattern and further emphasizing the features of the voice pattern, the magnitude of the similarity responds sensitively to the difference between the voice patterns. Therefore, even if a voice not to be recognized is input, it can be rejected with high accuracy without erroneous recognition.

FIG. 2 is a graph showing the results of an experiment conducted to examine the performance of the present invention. The recognition target word is 20 personal names, and each word is uttered three times to create a standard pattern. At the time of recognition, 40 words are uttered twice each including the 20 words to be recognized and the 20 words not to be recognized. A recognition experiment was performed. There are eight speakers, seven men and one woman. In the graph of FIG. 2, the horizontal axis represents the recognition rate and the vertical axis represents the rejection rate when the similarity threshold used for reject determination is changed. Here, the recognition rate and the rejection rate are defined as follows.

In the graph of FIG. 2, the solid line indicates the experimental result of the present invention, and the broken line indicates the experimental result when the similarity Si expressed by the above equation (10) is used as it is in the rejection processing.

As shown in the graph of FIG. 2, for example, when a recognition rate of 94.0% is realized, the rejection rate of only about 5% is obtained by the conventional method, whereas the rejection rate of about 35% is obtained according to the present invention. Rejection rate was obtained, and it became clear that rejection performance was greatly improved.

From the above experimental data, it can be seen that high reject performance can be obtained by using the differential pattern of the input voice for the similarity calculation of the reject processing. Therefore, it can be said that this method has a great effect in improving speech recognition performance.

The present invention is not limited to the embodiments described above. For example, although a standard pattern created by an orthogonalized time filter was used for both recognition processing and reject processing, other methods such as the so-called DP matching method were used for recognition processing, and the learned speech pattern was differentiated only for reject processing. A processed standard pattern may be used.

There are several variations of the coefficient of the differential processing filter. The point is that the similarity or difference between the standard pattern obtained by differentiating the learning voice pattern and the differential pattern obtained by differentiating the input voice pattern is obtained. A reject process is performed, and various modifications can be made.

Furthermore, the number of dimensions of the learning voice pattern is not particularly limited, and the present invention can be implemented with various modifications without departing from the gist thereof.

[Effects of the Invention] As described above, according to the present invention, both the learning voice pattern and the input voice pattern are differentiated, and the features of the voice pattern are further emphasized, so that the magnitude of the similarity is reduced. Since the difference becomes sensitive, it is possible to provide a speech recognition method that enables rejection with high accuracy without erroneously recognizing speech that is not recognized.

[Brief description of the drawings]

FIG. 1 is a schematic configuration diagram of a speech recognition apparatus configured by applying the speech recognition method according to the present invention, FIG. 2 is a graph showing the results of an experiment conducted to examine the performance of the present invention, and FIG. FIG. 7 is a diagram for explaining rejection processing when a voice not to be recognized is input. 1 ... Acoustic analysis unit, 2 ... Sound section detection unit, 3 ... Sample point extraction unit, 4 ... Sound pattern storage unit, 5 ... Orthogonalization time filter unit, 6 ... Standard pattern, 7 ... Recognition Similarity calculation unit, 8 ... smoothing pattern creation unit, 9 ... determination unit,
10 ... Rejection similarity calculation unit, 11 ... Differential pattern creation unit.

Claims (2)

(57) [Claims] A similarity or a difference is calculated between an input voice pattern obtained by analyzing an input voice and a standard pattern created based on a previously collected learning voice pattern. In the speech recognition method for recognizing the input speech, the standard pattern is created by using a plurality of filters that perform at least a smoothing process and a differentiation process on the learning speech pattern, and the standard obtained by the differentiation process. A speech recognition method comprising calculating a similarity or a difference between a pattern axis and a differential pattern obtained by differentiating an input speech pattern, and determining whether the speech is a recognition target speech. And calculating a similarity or a difference between an input voice pattern obtained by analyzing the input voice and a first standard pattern created based on a learning voice pattern collected in advance. Thus, in the voice recognition method for recognizing the input voice, a differential process is performed on the learning voice pattern to create a second standard pattern, and the generated second standard pattern and the input voice pattern are differentiated. A speech recognition method characterized by calculating a similarity or a difference between the differential pattern and the differential pattern obtained as a result, and determining whether or not the speech is a recognition target speech.