JPS6259999A - Unspecified speaker voice recognition equipment - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a speaker-independent speech recognition device, and particularly to a device having a function of adding standard patterns to a recognition dictionary for pattern matching with the aim of improving the recognition rate. The present invention relates to a specific speaker speech recognition device.

(Prior art) Conventional speech recognition methods targeting unspecified speakers include:
Input an analog audio signal, convert it A/D, and then convert it into some feature using a bandpass filter, LPC analysis, etc., calculate the similarity between the feature and the standard pattern, and find the maximum similarity. It becomes! Generally, a category attribute of a quasi-pattern is used as a recognition category.

The standard pattern is a pattern in which the feature amounts considered to be standard for the recognition target are prepared in advance for each category. How to create this standard pattern.

For each recognition target category, processing such as clustering and averaging is performed on the voice patterns, which are the feature quantities, of multiple speakers prepared in advance (generally about 1000 people), and multiple standard patterns per category ( In reality, it is common to create about 20 pieces. The voice pattern used to create the standard pattern will be referred to as a learning pattern hereinafter.

(Problems to be Solved by the Invention) The learning pattern used to create the standard pattern is made up of only a limited range of speakers, and the characteristic distribution of the limited speakers can be efficiently It is comprehensive. Therefore, when the learning pattern is recognized using the standard pattern, the possibility of misrecognition is extremely small, and because clustering is performed efficiently within each category, multiple standard patterns can be recognized in the feature space. Since the standard patterns are arranged evenly within the standard pattern, the frequency at which the degree of similarity is maximum among all the standard patterns when the learned pattern is correctly recognized is the same among the standard patterns.

However, when performing recognition targeting unspecified speakers, the number of speakers to be recognized far exceeds the number of speakers composing the learning pattern, and accordingly, the variation in feature distribution also increases within the learning pattern. This is large compared to the size of the dispersion. That is, the feature distribution of the learned pattern is different from the feature distribution of the unlearned pattern when recognition is performed for an unspecified speaker. Therefore, when recognition processing is performed on the unlearned pattern, an input pattern existing in the feature space outside the feature distribution range of the learned pattern will appear,
Such a pattern has the problem of causing an error P(m).

The present invention eliminates the drawbacks of the prior art described above, efficiently adds standard patterns to correctly recognize patterns that are incorrectly recognized among input patterns that have not been learned, and improves recognition performance. The present invention provides a speaker-independent speech recognition device that can perform the following tasks.

(Means for Solving the Problem) The present invention calculates the degree of similarity between an input voice pattern and a plurality of standard patterns created in advance, and calculates the category that the standard pattern having the maximum degree of similarity to the input voice pattern has. This relates to a speaker-independent speech recognition device that performs speech recognition by determining an attribute as a recognition category, and in order to solve the problems of the prior art, a first storage means,
The apparatus includes a category selection means, a second storage means, and an additional standard pattern creation means.

The first storage means corresponds to an erroneous recognition pattern storage unit in an embodiment described later, and the first storage means performs the speech recognition on a speech pattern (non-learning pattern) other than the speech pattern (learning pattern) used when creating the standard pattern. Record voice patterns that are misrecognized when processing is performed. The category selection means determines whether the number of misrecognitions in the speech recognition process, that is, the speech recognition process for the non-learning pattern is different from other LLI!
Select a category that is large compared to the target category.

The second storage means corresponds to the additional standard pattern creation learning pattern storage section in the embodiment, and stores the audio pattern belonging to the category selected by the category selection means, that is, the category with low recognition performance, from the first storage means. Enter and remember. The additional standard pattern creation means first calculates the degree of similarity between one pattern stored in the second storage means and all other patterns for each category, for all patterns in the same category. Next, the number of times the calculated similarity is greater than a predetermined threshold is counted for each category and each pattern. and a pattern corresponding to the average vector of the feature vector of the pattern counted the maximum number of times in the category and the feature vectors of all patterns in the same category whose similarity to the pattern is greater than the predetermined threshold value. Let be the standard pattern to be added in that category.

(Operation) Each technical means of the present invention operates as follows. First, the first
The storage means collects and stores patterns that are erroneously recognized in recognition processing performed on non-learning patterns. The category selection means selects a category for which recognition performance is low, that is, a category for which additional m quasi-patterns should be created, based on the misrecognition rate for each category in the recognition process of unlearned patterns.

The second storage means receives and stores a pattern having the same category attribute as the category selected by the category selection means from the first storage means. The additional standard pattern creation means performs the following operations for each category selected by the category selection means to create additional standard patterns. That is, a pattern corresponding to a certain category is received from the second storage means, and a pattern belonging to the category is received.
, and find the similarity between this pattern and all other patterns. Then, the number of times the obtained similarity is greater than a predetermined threshold is determined. This operation is performed for all patterns within the category.

Then, the pattern with the maximum number of times is detected, and a pattern corresponding to the average vector of the feature vectors of all patterns whose similarity to the pattern is greater than the threshold value and the feature vector of the pattern with the maximum number of times is found. Make the pattern an additional standard pattern in the category.

Therefore, recognition performance can be improved by minimizing the increase in the number of standard patterns.

(Embodiment) FIG. 1 is a block diagram showing the configuration of a speaker-independent speech recognition device according to an embodiment of the present invention, and FIG. 2 is a flowchart showing the operation of the device. As shown in FIG. 1, the speech recognition device of this embodiment includes a similarity calculation section 1, a control section 2, a standard pattern storage section 3, an erroneous recognition pattern storage section 4, a category selection section 5, and a learning section for creating additional standard patterns. It is composed of a pattern storage section 6 and an additional standard pattern creation section 7.

To explain the operation, the control section 2 outputs a recognition start command signal a to the similarity calculation section 1. The similarity calculation unit 1 receives this L& recognition start command signal, and sequentially inputs the input pattern b from the input time (step ■ in Fig. 2).
. At this time, the name of the 5 correct answer categories is also input at the same time. In addition, the input cover turn b here is a signal obtained by A/D converting the input audio signal and then converting it into a predetermined feature amount.

Then, the similarity calculation unit 1 calculates the similarity between the input cover pattern b and all the standard patterns stored in the standard pattern storage unit 3 (step ■), and outputs the recognition result C to the control unit 2. and an operation of outputting the misrecognition pattern d to the misrecognition pattern storage unit 4 only in the case of misrecognition (
Steps (2) and (2) are repeated until the input of input pattern b to be recognized is completed (step (2)). In order to confirm the effects of the present invention, it is desirable that the number of input patterns is 10,000 or more. After the input of the input cover turn b is completed, the similarity calculation unit 1 outputs a recognition completion signal e to the control unit 2. The control unit 2 calculates the false recognition rate for each recognition target category from the recognition result C for each speech pattern input from the similarity calculation unit 1, and calculates the recognition target category at the time when the recognition end signal e is input from the similarity calculation unit 1. The misrecognition rate f for each is output to the category selection section 5.

After inputting the false recognition rate f from the control unit 2, the category selection unit 5 calculates the name of the category and the number of categories whose false recognition rate is large compared to the false recognition rate of other recognition target categories (step ■). . In this embodiment, a category whose misrecognition rate is approximately 3% or more higher than the average value of the misrecognition rates of all recognition target categories is selected as a category for which an additional standard pattern should be created. The category selection unit 5 performs the following operations by setting one category among all the selected categories as an additional category name (step ①).

Input the misrecognized patterns one by one from the misrecognition pattern storage unit 4 (Step ■), and create additional standard patterns only for the patterns whose additional category names match the power category names of the input patterns (Step ■). The pattern input is output to the learning pattern storage section 6 (step [phase]), and it is determined whether the pattern input is completed (step 2). After the pattern input is completed, the category selection section 5 inputs an additional standard pattern creation start command signal g to the additional standard pattern creation section 7.

The additional #A quasi-pattern creating section 7 creates an additional standard pattern by a method described in detail below.

The number of learning patterns for creating additional standard patterns stored in the learning pattern storage unit 6 for creating additional standard patterns is GP.
NO, the number of feature dimensions of the learning pattern for creating an additional standard pattern is expressed by DNO, and the feature vector of the learning pattern for creating each additional standard pattern is expressed by the degree of similarity between each pattern (the concept of expressing the degree of similarity is the angle between vectors, Although urban area distance, Euclidean distance, etc. can be considered, in this explanation, Euclidean distance will be used as an example.) is defined by Euclidean distance. Here, it is assumed that the larger the Euclidean distance, the smaller the similarity, and the smaller the Euclidean distance, the larger the similarity. Then, the Euclidean distance is calculated between the learning pattern for creating an additional standard pattern and all other learning patterns for creating an additional standard pattern, and the Euclidean distance is smaller than a predetermined threshold RTH. Count the number of times (step @)
. The above process is performed for all additional standard pattern creation learning patterns, and the Euclidean distance is set to the threshold R.
The number of times smaller than TH is RPH, (i = 1~GPN
O). When the above operation is performed for all additional standard pattern creation learning patterns (step ■), the above RP among all additional standard pattern creation learning patterns
For the pattern with the maximum +, calculate an average vector between this pattern and all patterns for which the Euclidean distance to this pattern is smaller than the threshold RTH among other learning patterns for creating additional standard patterns. O), outputs the average vector as an additional standard pattern h to the standard pattern storage section 3 (step ■), and outputs an additional standard pattern creation end signal]- to the category selection section 5.
Output to.

The category selection section 5 outputs an additional standard pattern creation start command signal g to the additional standard pattern creation section 7 in order to carry out the above operations for all the categories for which additional standard patterns are to be created.According to this command, the additional standard pattern creation section 7 Perform the above additional standard pattern creation operation (Step O)
.

FIG. 3 conceptually represents the distribution on the feature space of learning patterns for creating additional standard patterns for one category. The operation of the additional standard pattern creating section 7 will be qualitatively explained below using FIG.

In FIG. 3, if the learning pattern for creating an additional standard pattern is A-L, the pattern in which the number of times the Euclidean distance D+ is smaller than the threshold RTH is maximum is A, and the Euclidean distance with A is the threshold The patterns that are smaller than RTH are B, C, D, E, and F, and A
, B, C, D, E, F, the average vector of six patterns in total is calculated and set as an additional standard pattern S.

In Figure 3, patterns such as G, J, and L are patterns that are located in a feature space that is far away from other erroneously recognized patterns among the erroneously recognized patterns, and these patterns are excluded from the above averaging target. As a result, additional standard patterns can be created using only patterns located in the feature space that frequently appear among misrecognized patterns, which is expected to minimize the increase in the number of standard patterns and efficiently improve recognition performance. can.

(Effects of the Invention) As explained in detail above, in the present invention, for categories where misrecognition is likely to occur, only patterns that appear frequently in the feature space among all misrecognition patterns are averaged and added. Since the standard pattern is used, the increase in the standard pattern can be minimized and recognition performance can be efficiently improved.

[BRIEF DESCRIPTION OF THE DRAWINGS] FIG. 1 is a block diagram of an embodiment of a speaker-independent speech recognition device according to the present invention, FIG. 2 is a flowchart showing the operation of the embodiment, and FIG. 3 is a block diagram of an embodiment of the speaker-independent speech recognition device according to the present invention. FIG. 3 is an explanatory diagram of an additional standard pattern creation method. 1... Similarity calculation unit, 2... Control unit, 3...
Standard pattern storage unit, 4... Erroneous recognition pattern storage unit, 5... Category selection unit, 6... Learning pattern storage unit for creating additional standard patterns, 7
...Additional standard pattern creation department.

Claims (1)

[Claims] Calculating the degree of similarity between the input voice pattern and a plurality of standard patterns created in advance, and determining the category attribute of the standard pattern that has the maximum degree of similarity to the input voice pattern as the recognition category. In a speaker-independent speech recognition device that performs speech recognition, a speech pattern that is erroneously recognized when the speech recognition process is performed on a speech pattern other than the speech pattern used when creating the standard pattern is stored. 1 storage means; a category selection means for selecting a category for which the number of incorrect recognitions in the speech recognition process is large compared to other recognition target categories; a second storage means for inputting and storing voice patterns from the first storage means; and one for each category stored in the second storage means.
The similarity between this pattern and all other patterns is calculated for all patterns in the same category, and the number of times the calculated similarity is greater than a predetermined threshold is calculated for each category.
The average of the feature vector of the pattern counted for each pattern with the maximum number of counts in the category and the feature vectors of all patterns in the same category whose similarity to the pattern is greater than the predetermined threshold value. A speaker-independent speech recognition device comprising additional standard pattern creation means for creating a pattern corresponding to the vector as a standard pattern to be added in the category.