JPH0318983A - Pattern collating system - Google Patents

Abstract

PURPOSE:To execute the category classification by the same distance measure even in the case lack of uniformity exists in a learning sample by synthesizing a vector to be referred to an unknown input vector by taking the accuracy of a feature vector into consideration. CONSTITUTION:A phoneme recognizing part 3 is provided with a reference vector synthesizing part 4 and a distance calculating part 5, and when a feature vector for representing a large category (i) and a feature vector for representing a small category (j) belonging to the large category (i) are denoted as Yi and Zij, respectively, a reference vector C to be compared with and referred to an inputted unknown vector X is synthesized by an expression of C = (1 - Wij)Yi +Wij Zij, and a distance between X and C is calculated. Wij is a constant of a range of 0<=Wij<=1 and an index for showing the reliability of Zij, and when Zij is generated by learning data of a sufficient number, a large value is taken. In such a way, even in the case lack of uniformity of reliability of the category caused by a difference of the number of learning samples exists, the category classification can be executed by the same distance measure or a similar measure.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a pattern matching method in a pattern matching section of a voice recognition or character recognition device.

[4] In vector quantization and category classification problems, the relationship between the number of categories and the number of learning samples is always an issue. In other words, the larger the number of categories, the more detailed classification can be made, but conversely, the number of learning samples per category becomes smaller, so the indicators (representative vectors, standard patterns, etc.) that express category-likeness become less accurate, and unknown input On the other hand, if the number of categories is reduced, the index representing category-likeness becomes statistically reliable, but different items may not be in the same category. There were drawbacks such as increased quantization distortion and increased quantization distortion.

Fuzzy vector quantization ("Normalization of spectrograms using fuzzy vector quantization" Journal of the Acoustical Society of Japan 45)
Vol. 2, No. 2 (1989)) improves on this drawback and can reduce quantization distortion even with a small number of categories.

However, when detailed classification became necessary, the above drawbacks were still not analyzed. Furthermore, conventional techniques have had the disadvantage of being difficult to deal with variations in the number of learning samples. For example, consider the case where phonemes are registered as standard patterns in a speech recognition device for a specific speaker. Since it is nearly impossible for humans to utter words in units of phonemes, for example, they utter words in units of words, which are then cut out in units of phonemes to create a standard pattern. The problem here is that the frequency distribution of phonemes is uneven. For example, a situation may occur in which 100 pieces of data for /a/ are obtained, but only two pieces of data for /p/ are obtained. As a result, the standard pattern for /a/ is statistically reliable enough, but a highly accurate standard pattern cannot be expected for /p/. Further, regarding /p/, recognition using HMM, Bayesian judgment, Mahalanobis distance, etc. becomes impossible. Furthermore, in extreme cases, there may be cases where there is no utterance for a prepared category (phoneme).

As described above, if the training samples are uneven, category classification may not be possible using the same distance measure.

且-□拵The present invention was made in view of the above-mentioned circumstances,
The purpose of this method is to provide a pattern matching method that enables category classification using the same distance measure or similarity measure even when there are variations in the reliability of categories due to differences in the number of training samples.

(ii) In order to achieve the above object, the present invention uses a pattern matching method that matches a human-generated unknown vector X with a feature vector representing a category, in which each category is classified into M large categories. , where Yi is the feature vector representing the large category i, the large category i is further classified into N(1) small categories, and Zij is the feature vector representing the small category j belonging to the large category i. , when calculating the degree to which the unknown vector X belongs to the small category j within the large category i, or the distance between X and the small category j within the large category i.

C=(I Wij) Yi ten Wij Zij
This method is characterized in that the degree of belonging or the distance is calculated with reference to a composite vector C satisfying O≦Wij≦1. Hereinafter, the present invention will be explained based on examples.

FIG. 1 is a system configuration diagram for explaining an embodiment in which the pattern matching method of the present invention is applied to a pattern matching section in specific speaker speech recognition. In the figure, 1 is a microphone;
2, the feature sequence conversion section 53 is a phoneme recognition section.

The audio waveform input from the microphone 1 is processed by the feature series converter 2
is converted into a time series of feature vectors.

Various types of feature vectors and means for converting them are known that are effective for speech recognition. For example, 12K) [z.

After A/D conversion with 12 bits, window length is 256 pain
t and a shift width of 128 paint to find the 14th linear prediction coefficient.

Thereafter, the sound '-h J recognition section 3 performs phoneme recognition on the feature vector X. Here, X may be a vector for one frame or a vector grouped in units of several frames.

In phoneme recognition, in order to avoid the effects of articulatory combination, it is desirable to prepare different standard patterns for each preceding and following phonetic environment. For example, it is preferable to prepare five types of phoneme standard patterns for /ni/ for each subsequent vowel. However, preparing standard patterns for all phonetic environments, that is, having speakers register them, would result in a huge number of utterances, which is not realistic.

FIG. 2 is a diagram showing the configuration of the phoneme recognition unit, in which 4 is the reference vector synthesis unit, 5 is the distance calculation unit, and 6 is the reliability level Wij.
, 7 is a standard pattern Zij, and 8 is a standard pattern Yi. The standard patterns are a standard pattern Yi (1≦i≦M, M is the number of phonemes) created for each phoneme, and a standard pattern Zij (1≦j≦N
Two types are prepared: (i) and N (i) is the number of environments of phoneme i). For example, in the above-mentioned example of /ni/, there are Yi (which may be plural) created without considering the phonetic environment and Zij, which are prepared in five types for each subsequent vowel. (
N(i)=5) Then, compare the input unknown vector
The distance between X and C is calculated by combining them using the formula (1). In equation (1), Wij is a constant in the range of 0≦Wij≦1. Wij is an index representing the reliability of Zij.

When Zij is created with a sufficient number of learning data, it takes a large value. Conversely, if Zij is less reliable, Yij is more trusted and C is synthesized.

For example, in the example of /ni/ mentioned above, when there was less training data for /ku/, the whole of /ka/ to /ko/ was created with /ni/'s [141! Equation (1) is set so that pattern Yij is trusted and C cape Yij is obtained.

The distance d between X and C may be calculated using, for example, Euclidean distance as follows: a=lx-all'' (2).

As is clear from the above explanation, according to the present invention, the reference vector synthesis unit synthesizes the vector C to be referenced by the unknown input vector There is. For this reason, if Zij is created from a very small number of learning data and the accuracy of Zjj is low, by setting the 1 weight Wij small, C can be synthesized by rough approximation by Yi representing the large category i. can do.

Conversely, Zi representing the small category j within the large category i
If j is created from many learning data, Wi
By setting j large, a precise reference vector can be synthesized.

Further, even if there is variation in reliability of each ZJJ due to variation in the number of learning data for each small category, reference vectors can be synthesized using the same synthetic force method according to the present invention.

[Brief explanation of drawings]

FIG. 1 is a system configuration diagram for explaining an embodiment in which the pattern matching method of the present invention is applied to a pattern matching unit in specific speaker speech recognition, and FIG. 2 is a configuration diagram of a phoneme recognition unit. 1... Microphone, 2... Feature sequence conversion unit, 3... Phoneme recognition unit, 4... Reference vector synthesis unit, 5... Distance calculation unit, 6... Reliability, 7, 8...・4i1$ pattern.

Claims (1)

[Claims] 1. In a pattern matching method that matches an input unknown vector X with a feature vector representing a category,
Each category is divided into M large categories,
Let Yi be the feature vector representing the large category i, the large category i is further classified into N(i) small categories, and let Zij be the feature vector representing the small category j belonging to the large category i, When calculating the degree to which the unknown vector X belongs to the small category j within the large category i, or the distance between X and the small category j within the large category i, C=(1-Wij)Yi+WijZ1j0≦Wij≦ 1
A pattern matching method characterized in that the degree of attribution or distance is calculated by referring to a composite vector C.