JPH04307599A - Word voice recognizer - 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]

0001

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a word speech recognition device for recognizing input words by voice.

0002

[Prior Art] A conventional word speech recognition device extracts the speech features of the entire input word, compares the features with each of the pre-registered features of the word to be recognized, and compares the features with each of the features of the word to be recognized. The recognition results were words that were considered to be similar to each other based on certain criteria.

FIG. 3 shows a schematic diagram of such a conventional device.

In FIG. 3, numeral 15 is a voice input means, 16 is a voice feature extraction means for extracting voice features from the voice input by 15, and 17 is a voice feature extraction means for processing the features extracted by 16 as a newly registered word. 18 is a registered word feature storage means for storing the feature string from 16 as a recognition target; 19 is a registering word feature storage means for storing the voice features generated by 16 in 18; Each pattern matching means for comparing features of words is schematically shown.

[0005]

[Problems to be Solved by the Invention] However, since word speech recognition devices such as this conventional example treat the entire feature of a word as a recognition target, variations due to various factors within and between speakers are tolerated. It has to be possible.

[0006] In this case, the simplest method that can be considered is to extract more general feature quantities based on data of several hundred people and use them for recognition. However, with this general method that uses only feature amounts, it is difficult to sufficiently compensate for variations between speakers, and a high recognition rate cannot be achieved. In order to improve the recognition rate with such a method, speaker adaptation processing is required.

[0007] An extreme and reliable processing method for speaker adaptation is a method in which all words to be recognized are registered in the utterances of each user.
However, in such a method, it requires a very large amount of work to additionally register new words. It also requires a very large storage device to store the data.

For example, for 100 users, 100
Assume that a device that recognizes words is needed, and further assume that it is necessary to utter one word three times in order to register data for each person. Then, in order to register this data, a total of 30,000 utterances would have to be processed, and a total of 10,000 words of data would need to be stored. This is a very strong restriction in practical use and is completely unacceptable.

This is the problem that the present invention aims to solve, and its purpose is to use general word feature data that is more independent of speakers, and to extremely easily perform speaker-adaptive processing. The goal is to realize a word speech recognition device that can perform this.

0010

Means for Solving the Problems FIG. 1 is a schematic diagram of the concept of the present invention.

As shown in the figure, the present invention provides a word speech recognition device that includes at least a) a speech input means; 2. For the feature parameter strings extracted by c) and b), determine the degree of similarity with at least the phoneme characteristics of each of the vowels a, i, u, e, o, and N, and feature similarity vector extracting means 3 that expresses as a feature similarity vector of at least six dimensions, and whether to start recognition processing or registration processing for the feature similarity vectors extracted by e) and c). A recognition/registration recognition operation control means 4 for selecting a recognition/registration recognition operation control means 4, a registered word feature similarity vector storage means 5 for storing the feature similarity vector selected by e) and e) and extracted by c) as a registered word, and f) , a pattern matching means 6 that performs a pattern matching operation between the feature similarity vector selected by d) and extracted by c) and the registered word feature similarity vector of e); This is a word speech recognition device.

0012

[Example] The essence of the present invention is that the characteristics of all phonemes are divided into at least six phonemes /a/, /i/, /u/, /e/, /
This is expressed by expanding the feature vectors of o/ and /N/ as basis vectors.

FIG. 1 is a conceptual diagram of an embodiment of the present invention, and FIG. 2(a) schematically shows the basic principle of the present invention.
The present invention will be explained in detail based on b).

FIGS. 2(a) and 2(b) show /a/, /i/, /u/, /e/, /o/, and /N, which are the constituent elements of the present invention.
Two of the feature similarity vectors for phonemes such as / are virtually extracted, and the basic principle of the present invention will be explained using the two-dimensional feature similarity vectors. Figure 2
(a) shows a feature vector space for a certain speaker A, and it is assumed that number 7 in the figure indicates the feature vector of /a/, and number 8 similarly indicates the feature vector of /i/.
Also, to simplify the explanation, it is assumed that these two vectors are orthogonal. Similarly, FIG. 2(b) shows the feature vector space for speaker B, and number 11 in the figure
Let it be assumed that the number 12 indicates the feature vector of /a/ and the number 12 indicates the feature vector of /i/.

[0015] Here, two hypothetical phonemes /x/ and /y/
Suppose that they are expressed as numbers 9, 10, 13, and 14 in this vector space for speakers A and B, respectively. The vectors corresponding to /x/ and /y/ are written as components in the feature vector space of speakers A and B, respectively, and are /x/=(4, 3) and /y/=(3, 4).

However, if speaker B's phoneme /x/ is expressed in speaker A's feature vector space, this feature will be closer to speaker A's phoneme /y/ than to speaker A's phoneme /x/. In other words, if speaker A's features are more standard features and we try to recognize the utterances of all speakers using that feature data, in speaker B's utterances the phoneme /x/ will change to the phoneme /x/. It will be mistakenly recognized as y/.

In other words, in the feature space formed by the base feature vectors for each speaker, the phonemes of each speaker are considered to have equivalent features. but,
If we try to express it in the feature space formed by the base feature vectors of other speakers, the feature vectors will differ slightly due to individual differences between speakers, and therefore will be expressed close to the features of different phonemes. What you say happens.

In order to solve this problem, the present invention uses a set of feature vectors /a/, /i/ based on the speaker's own utterances.
, /u/, /e/, /o/, /N/, etc., are used as base feature vectors to create speaker-independent words by expressing arbitrary phonemes as feature similarity vectors. This makes it possible to extract audio features. Here, as a minimum case, adopting the feature vectors of these six phonemes as the base feature vector may have some problems in terms of accuracy in expressing arbitrary feature vectors. However, in speaker adaptation processing, the data input by the speaker must be easy to utter, and the adaptation processing itself must be easy.

/a/, /i/, /u/, in the present invention
/e/, /o/, /N/, etc. are all extremely easy to utter and have the characteristic of being constant in time.

[0020] Therefore, in the adaptive processing, operations such as cutting out specific data are not necessary.

[0021] Furthermore, by using the average value or the like within the utterance time, it is possible to reduce the influence of noise during data input. Even if these features are somewhat less accurate in expressing the feature vector space, they can sufficiently compensate for the drawbacks. Of course, by adding some base feature vectors to these, the accuracy of feature expression can be increased. However, this also increases processing for data collection and the like. Therefore, the number of base feature vectors should be considered as the optimum point considering the required accuracy of feature expression and increase in processing amount.

FIG. 1 schematically shows an example of the actual configuration of the present invention. First, word speech is input by the speech input means 1. For example, it is input as A/D converted digital data. The input data is the output of FFT bandpass filter or LPC
The feature extraction means 2 extracts the feature as a cepstral coefficient or the like. Then, the extracted features are
The feature similarity vector extraction means 3 converts it into a feature similarity vector. In the registration operation, the recognition/registration operation control means 4 registers the obtained feature similarity vector as a recognition target in the registered word similarity vector storage means,
In the recognition operation, the pattern matching means 6 is activated to perform pattern matching between the obtained feature similarity vector and the registered word data in the registered word feature similarity vector storage means.

[0023]

According to the present invention, the speaker adaptation process simply adapts the new speaker to /a/, /i/, /u/, /e/, /o/,
All you have to do is ask them to say /N/ etc.

[0024] Furthermore, since the registered word data does not depend on the speaker, it can be used in common among speakers.
There is no need for each speaker to register the words again. Therefore, the data for each word can be handled as if it were data from a normal character dictionary.

[0025] Furthermore, this adaptive data is information that depends only on the individual. Therefore, if you already have this information, it is possible to use the speech recognition device without performing the above adaptation procedure. It is also possible to authenticate individuals using it. Furthermore, since this data is very small, even in a speech recognition system for a large number of people, the storage capacity required to hold the data is also very small.

The method according to the invention can be easily applied to continuous speech recognition devices.

[Brief explanation of the drawing]

FIG. 1 is a conceptual diagram of a word speech recognition device according to the present invention.

FIG. 2(a) is a diagram showing an example of expression of phoneme features in a feature vector space for a certain speaker for explaining the basic principle of the present invention. (b) is a diagram showing an example of expression of phoneme features in a feature vector space of another speaker.

FIG. 3 is a conceptual diagram of a conventional word speech recognition device.

[Explanation of symbols]

1: Voice input means 2: Feature extraction means 3: Feature similarity vector extraction means 4: Recognition/registration operation control means 5: Registered word feature similarity vector storage means 6: Pattern matching means 7: Base phoneme feature vector 8: Base Phoneme feature vector 9: Phoneme example 10: Phoneme example 11: Base feature vector 12: Base feature vector 13: Phoneme example 14: Phoneme example 15: Voice input means 16: Voice feature extraction means 17: Recognition registration operation control means 18: Registration Word feature storage means 19: pattern matching means

Claims (1)

[Claims] [Claim 1] A word speech recognition device, which device comprises at least a), a voice input means, b), a feature extraction means for extracting a feature parameter string from the voice input in a), and c), a). ), the degree of similarity with the phoneme characteristics of each of the vowels a, i, u, e, o, and N is determined, and expressed as a feature similarity vector of at least six dimensions. A recognition/registration recognition operation control means for selecting whether to start recognition processing or registration processing for the feature similarity vectors extracted in d) and iii); )
, registered word feature similarity vector storage means for storing the feature similarity vectors selected by , d) and extracted by iii) as registered words; A word speech recognition device comprising: a pattern matching means for performing a pattern matching operation between the degree vector and the registered word feature similarity vector of (e);