JPS6075891A - Phoneme segmentation - 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 phoneme segmentation method for performing phoneme recognition of input speech and recognizing syllables, words, sentences, etc. using the results.

(Structure of conventional example and its problems) FIG. 1 schematically shows the functions of an apparatus that implements a conventional segmentation method. This conventional example will be explained below with reference to FIG. In FIG. 1, 1 is a parameter extraction unit for segmentation, 2 is a reading of a dictionary item Dn in a word dictionary, 5 is a segmentation rule part for each phoneme, and 6 is a memory for phoneme standard patterns. 7 indicates a distance calculation unit that calculates the distance from the standard y-tan for each frame. The phoneme standard pattern represents the average value of each phoneme in various parameters, and is created in advance through preliminary experiments. N ・Q rays in a certain frame are cl, c2...c
Let N be the average value of those parameters for a certain phoneme .

If the target word to be matched with the input speech is 9, the dictionary item Dn in the word dictionary is read by the dictionary item reading unit 2,
In accordance with the phoneme sequence of Dn, the phoneme boundary candidate position determining unit 4
Segmentation is performed from n to Pnm'-i. Phoneme segmentation is performed by segmentation rule section 5.
dx ((-++
Changes in C2...CN)/? Performed as a parameter. However, in the above conventional example, since the standard pattern has only the average value of the parameters, the degree of variation in the parameter values is not reflected in the distance calculation using formula 0, and segmentation errors occur for phonemes with large variations. There were drawbacks.

(Objective of the Invention) The present invention eliminates the drawbacks of the above-mentioned conventional example, and aims to reduce speech recognition errors caused by segmentation errors.

(Structure of the Invention) In order to achieve the above object, the present invention compares the parameter time series obtained from the input speech with the phoneme standard, turn and word dictionary, and performs segmentation of the phonemes constituting each dictionary item. By using the probability density value indicating the value of the e-parameter of a phoneme, the effect of reducing segmentation errors is obtained.

(Description of Embodiment) An embodiment of the present invention will be described below with reference to the drawings. In FIG. 2, numerals 1 to 5 are the same as in the conventional example shown in FIG. Reference numeral 8 denotes a phoneme standard button unit that stores phoneme standard numbers consisting of average values and covariance values for each phoneme, and 9 is a probability density calculation unit for each phoneme.

In FIG. 2, a nosolameter extraction unit 1 analyzes input speech frame by frame and extracts parameters, and a probability density calculation unit 9 calculates the probability density that the parameters are generated from each phoneme. Next, the dictionary item reading unit 2 reads the recognition target word Dn from the word dictionary unit 3, and detects the boundary using the probability density as a parameter according to the segmentation rule 5 for each constituent phoneme.

Here, the calculation in the probability density calculation section 9 is performed using the following equation 0. A vector C represents n parameters in a certain frame. If we express the average value of a phoneme X in the parameters of a certain phoneme X stored in the phoneme standard/tan section 8 and the covariance value of X1 as Σ, then the probability density φx (C) of the vector C for the phoneme X is ...Represented by ■. However, the distribution shape is assumed to be Gaussian distribution.

In this embodiment, by using changes in the probability density for each phoneme in the dictionary phoneme series, boundary detection can be performed taking into account the degree of variation in parameters, which has the advantage of improving segmentation and accuracy.

(Effects of the Invention) The present invention has the above configuration, and provides the following effects. By using the probability density value in which each phoneme indicates the value of its parameter for segmentation of the phonemes constituting each dictionary entry, there is an advantage that segmentation accuracy for phonemes with large variations can be improved.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the functions of an apparatus for implementing a conventional phoneme segmentation method, and FIG. 2 is a diagram showing the functions of an apparatus for implementing a phoneme segmentation method according to an embodiment of the present invention. 1... Rigram meter extraction section, 2... Dictionary item reading section,
3... Word dictionary section, 4... Phoneme boundary candidate determination section, 5
...Segmentation rules part, 8... Phoneme standard? Tan part, 9... Probability density calculation part. '45. i;. Figure 1 Figure 2

Claims (1)

[Claims] The parameters of the input speech were extracted, and the parameters were compared with a standard phoneme pattern to calculate the probability density that the above-mentioned mother makeup is generated from each phoneme, and the word to be recognized was expressed as a symbol string for each phoneme. A phoneme segmentation method characterized in that each dictionary item in a word dictionary is matched with input speech, and phoneme boundaries are detected using the value of the probability density as a parameter according to a dictionary phoneme sequence constituting each dictionary item.