JPS6068393A - Discrimination of phoneme - 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 the Invention The present invention relates to a phoneme recognition method in a phoneme recognition method characterized by performing phoneme recognition.

Structure of conventional examples and their problems In recent years, research and development on speech recognition targeting unspecified speakers and multiple languages has become active.

Speech recognition methods are characterized by the fact that they are not susceptible to speaker-specific variations such as differences in accents, and that they convert speech signals into phoneme sequences, symbols that contain a small amount of information and are compatible with linguistics. This method is suitable for recognition of a wide variety of speakers and many words, since the capacity of the word dictionary may be small, and the contents of the word dictionary can be easily created and changed. The important point in this method is to perform phoneme recognition accurately. Aside from vowel recognition, consonant recognition has traditionally been considered a technically difficult problem. Therefore, I will mainly focus on consonants here.

Many phonetic studies have been conducted on the characteristics of individual consonants. However, there are not many conventional examples of so-called automatic recognition, a method of segmenting a speech signal and detecting all consonant intervals to recognize phonemes. Here, we will take up all the conventional techniques that were previously applied by a group including the present applicant, and list the problems.

A conventional method for detecting phoneme sections will be explained with reference to FIG. The following three methods are used together.

B. Perform speech/unvoiced determination and determine the unvoiced section as an all-consonant section. As shown in Figure 1 (IL), the audio is 10 m'ieC.
Divide into frames of about ', and voice/
Performs silent judgment. Then, a section in which unvoiced frames are continuous is defined as a consonant section. This method is effective for detecting voiceless consonants.

11 opening HGO-68393 (2) Each frame of the mouth and speech interval is divided into 6 vowels and a nasal sound (/m/.

The phoneme with the highest degree of similarity is selected as the recognition result for that frame. Then, a section in which frames recognized as nasal sounds are consecutive is defined as an all-consonant section (FIG. 1(b)). This is a segmentation method that takes advantage of the fact that the spectrum of voiced consonants is closer to nasal sounds than to vowels.

C. Detect a power dip (power delag) from the temporal movement of voice power, and define the dip section as a consonant section (Fig. 1 (C)). This is a segmentation method that takes advantage of the fact that the power of consonants is smaller than the power of vowels.

In the conventional method, all consonants are recognized in the consonant section obtained in this manner as follows. This will be explained with reference to FIG. The degree of similarity between each consonant and the standard pattern is calculated for each frame, targeting all frames L/- of the consonant section. In the figure, the similarity f/a for phoneme A, the similarity 'fxlb for phoneme B, and the similarity k1 for phoneme c. It is expressed as c. The similarity is (formula 1) 7! j=-(X-μ,)T・Σ
51.(x])-Kj (Formula 1) where j is the phoneme name (a, b, c...), and Kj is a constant that depends on the phoneme j. Also, χ is the input feature parameter (Lp
c kegstrum coefficient) vector, the edge is the mean value vector (
standard pattern), Σ is the covariance matrix (standard pattern).

Then, calculate the sum of similarities of phonemes A, B, C, etc. for all frames, respectively, and calculate L, Lb, Lo...
Then, it becomes. In this way, the phoneme ' with the maximum similarity is determined to be the recognized phoneme.

The problem with the conventional example described above is in the second half of the phoneme discrimination part, in that after an interval is determined by segmentation, similarity calculation is performed for each frame for the entire interval.

In other words, the entire consonant interval is assumed to be static throughout time, and the entire consonant interval is treated equally.

Page 6However, apart from vowels, the characteristic parameters of consonants and semi-vowels change over time within an interval, and the characteristics of each phoneme are found in the form of these changes. The characteristic part (characteristic part) differs depending on the type of consonant or semivowel. For example, in voiced and voiceless plosives, the features for identifying phonemes are concentrated near the plosive, and for nasals, the features for identifying phonemes are present in the transition to the following vowel.
Original sounds and semi-vowels have characteristics in the movement of parameters throughout the phoneme interval.

Therefore, an effective method for distinguishing between consonants and semi-vowels is to accurately extract the characteristic parts for discriminating each phoneme, and perform phoneme discrimination by focusing on the temporal movement of parameters in the characteristic parts. In the conventional example, such consideration has not been taken.

Purpose of the Invention The present invention solves the above-mentioned drawbacks of the prior art, and aims to accurately extract all the characteristic parts of a phoneme and match them with a standard phoneme pattern, including the temporal movement of parameters in the characteristic parts. From page 7 onwards, the purpose of this paper is to provide a means for identifying all phonemes with high accuracy.

Structure of the Invention The present invention achieves the above object, and uses a standard pattern (hereinafter referred to as a phoneme standard pattern) created for each phoneme in a part of the phoneme interval that well expresses the characteristics of the phoneme (hereinafter referred to as the characteristic part). ) and a standard pattern (hereinafter referred to as the surrounding information standard pattern) created for the surrounding information of the feature part regarding the phoneme group to be identified, and perform segmentation of the input speech to identify the feature part in the phoneme interval. After finding a candidate section, pattern matching is performed by applying the phoneme standard pattern and surrounding information standard pattern to each point in the feature candidate section, and all effects of the surroundings of the feature are removed based on the degree of similarity with each phoneme. About the entire feature candidate section in the form of
The present invention provides a phoneme discrimination method characterized in that phonemes are discriminated by comparing all similarities within the characteristic part candidate section.

Description of examples JP-A-Sho GO-68393 (3) An embodiment of the present invention will be described below with reference to the drawings.

In this example, the features of each phoneme are accurately detected visually, a standard phoneme pattern is created in advance using a large amount of data, and the similarity of all phonemes (belonging to a phoneme group) is compared at once. A standard pattern for the surrounding information of the special part has been created for phonemes), and for phoneme discrimination, first set the candidate section of the characteristic part a little wider using the segmentation parameter, then set the candidate section of the feature part a little wider, and then By applying the standard pattern of each phoneme and the standard pattern of surrounding information while shifting them one frame at a time and calculating the degree of similarity, features can be accurately detected from candidate intervals and phonemes can be discriminated. be.

First, a method for creating a standard pattern will be explained. The phoneme standard pattern is created by extracting all the feature parameters for the feature parts visually extracted for each phoneme, and then calculating the average value and covariance matrix of a lot of data. All LPG cepstrum coefficients obtained by linear prediction analysis 9 pages (LPG analysis) are used as the characteristic parameters.

Let the number of feature parameters per frame be d,
Letting the number of frames of the feature part be j, the parameter sequence χ of the feature part is (j) (j) (j)! 1, x2-xd) (Equation 2). x(k) is the i-th LPG cepnutrum coefficient in the first frame of the feature.

A parameter series is extracted from a large amount of data, and the mean value vector L of each element and the covariance matrix Σ between the elements are used as a standard pattern.

(j) (J) (j) μm, μ2...μd) (Formula 3) where μ year) = f x (kゝ/M (M is the number of data) l z
, I) The covariance matrix is complex and will not be described here. As described above, the method of this embodiment is characterized in that all standard patterns are created using characteristic parameters of a plurality of frames, that is, taking into consideration the temporal movement of the parameters.

Next, a method for creating a standard pattern for surrounding information will be explained. One type of standard pattern of surrounding information is created for a group of phonemes whose similarities are compared with each other in phoneme discrimination. for example,
Voiced plosive group (/b/, /(1/.

/q/) and one for the nasal group (/m/, /n/
, /17/).

The standard pattern of surrounding information is a pattern of the nature of information surrounding a feature. Each phoneme group has properties common to that phoneme group. For example, the characteristic part of a voiced plosive is near the plosive, but there is always a buzz section of several frames before the plosive, and after the plosive, it quickly connects to a vowel. In voiceless plosives, there is a silent section before the characteristic part (near the plosive).

In nasal sounds, there is a constant nasal frame before the characteristic part (the part that transitions to the following vowel).

The standard pattern of surrounding information is a standard pattern of characteristics common to such phoneme groups.

FIG. 3 shows an example of a specific production method. The surrounding information section is determined by setting sections 11 and 11 of sufficient length before and after the characteristic portion (the shaded section in the figure). The length of the interval is set for each phoneme group. For the surrounding information section, the dxβ-dimensional parameter series is set for the entire section by dividing it into β frames and shifting it by one frame as shown in the figure. This operation is applied to all data belonging to the phoneme group, and a standard pattern for surrounding information is created in the same manner as in the case of creating a phoneme standard pattern. If this is done, the data of the characteristic part will also be mixed in, but this will not be a problem because the weight of stationary surrounding information is much greater.

Next, a method for recognizing phonemes will be explained. First, a consonant interval is determined using a segmentation parameter, and the rear end of the consonant interval is used as a reference point, and all candidate intervals of the characteristic part are set to be somewhat wider. Although any segmentation parameter may be used, in this embodiment, it is used mainly for power dips in the high and low ranges, and all the rising parts of the dips are used as reference points. When the power dip is difficult to detect, the voiced/unvoiced judgment result and nasality are used together.

Next, a complete description will be given of a method for detecting a feature part from a feature part candidate section and discriminating a phoneme. In the following explanation, for simplicity, it is assumed that the phoneme group is composed of two phonemes (phoneme 1 and phoneme 2). Even if the number of phonemes increases, the idea remains the same.

It is assumed that the characteristic part candidate section is tt1 to t2. Ishi, point 1 (1
, ≦t≦t2), let X be the unknown input vector (data to be determined). χ has the same format as (Equation 2). And the standard pattern (average value) of phoneme 1 tt
u4. Standard pattern (average value) of phoneme 2. Standard pattern (average value) of Woku 1 surrounding information. and heel phoneme 1. Phoneme 2
Let Σ be the covariance matrix common to all of the surrounding information.

Σ is created by averaging each covariance matrix.

The similarity (distance) between the unknown input and phoneme 1 at time t is calculated as B19. Then Ll, t = (X t/41)T・Σ4・(Xt−
Thou,)-(X:t-μ.)・Σ・Sentence,-μ. ) (Formula 4)
Similarly, if the distance to phoneme 2 is kLz, t, then 134-B2. t=(Xt102)T・X−1・(Xttt2
)-(Xt-μ.)・Σ-'-(Xt-μ6) (Equation 6
). The first term in these equations is the Mahalanobis distance for the phoneme, and the second term is the Mahalanobis distance for the surrounding information. Therefore, the meaning of these equations is that the similarity (distance) between the unknown input and the phoneme standard pattern at time t minus the distance to the surrounding information is set as the new distance to the phoneme. The calculations of (Formula 4) and (Formula 6) are performed for the period from t1 to t2, and Ll, t,
B2. Among t, the phoneme that is the smallest during this period is set as the recognized phoneme.

In reality, (Formula 4) and (Formula 6) can be expanded into a simple formula as follows (the derivation is omitted as it is not essential).

Ll, ta, ·jCt-β, (Formula 4)'L2. t=t
2・Jt-lB2 (Formula 6)'it,, L2+ IB,
, IB2'e is newly set as a standard pattern that includes surrounding information.

A conceptual explanation of the above method is given in FIG.

In the situation shown in Figure 4e), the consonant discrimination is performed in line 14.
,, ,,.

Consider the case where Assume that the anti-shock feature candidate section T is determined as t, to t2 in the true feature (shaded area) of this consonant. (b) il-1' phoneme 1. The temporal variation of the distance to phoneme 2 is shown by a solid line and a broken line, respectively. A, B, and C indicate positions where the distance is minimum. In the true characteristic portion (point B), the portion of phoneme 1 is smaller than that of phoneme 2, and this consonant should be determined as phoneme 1. However, within the feature candidate section automatically determined by the segmentation parameters, phoneme 2 is at its minimum at point A, so if this continues, it will likely be misclassified as phoneme 2. FIG. 4 (Q) shows the distance between the unknown input and the standard pattern of surrounding information, and the value increases near the true feature. This is because the standard pattern is created mainly using peripheral information. Figure 4 ((1) is the distance from the standard phoneme pattern that includes surrounding information,
It is equivalent to Φ) minus (c) e. In (d), the value at point B is smaller than that at point A, and this consonant is correctly identified as phoneme 1.

167°,.

In this manner, by using the method of the present invention, it is possible to accurately extract true features from the rough feature candidate sections determined by the segmentation parameters and to discriminate all phonemes.

In addition, although the explanation was made using the Mahalanobis distance above,
The idea is the same for other distances. For example, when using (Equation 1), use likelihood instead of distance,
It works if you use the local maximum value instead of the local minimum value. Further, although the above explanation has been made using consonants, a similar method can be applied to phonemes that change over time, such as semi-vowels.

FIG. S is a block diagram for implementing the method of the present invention.

This is a diagram. Reference numeral 1 denotes a feature parameter extraction unit, which performs all analysis of input speech and extracts feature parameters.

Here, LPG analysis is performed and LPG cepstrum coefficients are determined and sent to the similarity calculation section 2. Further, the feature parameter extracting section 1 collects high-frequency power and low-frequency power of the input voice and sends them to the segmentation section 4. The similarity calculation unit 2 calculates the distance between the input parameter and each standard pattern stored in the standard pattern storage unit 3 according to Japanese Patent Application Laid-open No. Sho GO-68393(5) for each frame. In addition to phoneme standard patterns and surrounding information standard patterns, the standard pattern storage unit 3 also stores voiced/unvoiced standard patterns and five vowel/nasal standard patterns used for segmentation.

In the segmentation 84, phoneme sections are determined based on the similarity information sent from the similarity calculation section 2 and the power information sent from the feature parameter extraction section 1. The phoneme discriminator 6 extracts all the feature parts based on the similarity between the phoneme interval and each phoneme, and the similarity of surrounding information, compares all the similarities of the phonemes in the feature parts, discriminates the phoneme, and outputs the result.

According to this example, an average of 76.1
% recognition rate was obtained. A similar evaluation using the conventional method would result in 72
．． It was 6%. Considering that some consonant groups are recognized in the conventional method, it can be seen that the effect of this embodiment is remarkable.

The data used was 21217 utterances made by a total of 2 men and women.
・ −・ In order to investigate the effect of using standard patterns of surrounding information, we conducted experiments using voiced plosives and nasals.

As a result, when the standard pattern of surrounding information was not used, the recognition rate was 7J7% for vocal plosives and 64.1% for nasal sounds, whereas when the standard pattern of surrounding information was used,
This improved to 74.7%, 7es, and 2%, respectively. In particular, a remarkable effect appears on nasal sounds. This is because the power delag of nasal sounds is unclear, so phoneme intervals cannot be extracted accurately. When the standard pattern 7 of surrounding information was introduced, even when the phoneme section was unclear, the characteristic part could be detected accurately and the recognition rate was improved, thereby verifying the effect of this example.

Effects of the Invention In short, the present invention provides a standard pattern created for each phoneme for a part of a phoneme interval that well expresses the characteristics of a phoneme (hereinafter referred to as a characteristic part), and a characteristic pattern for a group of phonemes to be identified. Prepare a standard pattern created for the surrounding information of the part, perform all the segmentation of the input speech, find the feature 18 in the phoneme interval, ... part candidate interval, and set it for each point in the feature part candidate interval. All pattern matching is performed by applying the phoneme standard pattern and the surrounding information standard pattern, and the similarity with each phoneme is calculated for the entire feature part candidate section in a form that removes the influence of the surroundings of all the features, and the feature part candidate section is The present invention provides a phoneme discrimination method that is characterized by discriminating phonemes by comparing the degree of similarity within them.B. Automatic segmentation of speech can be performed to discriminate phonemes with high accuracy.

By using the mouth, phoneme standard pattern, and surrounding information standard pattern, parts (characteristic parts) effective for phoneme discrimination can be fully automatically and accurately extracted and matched.

It has the following advantages.

[Brief explanation of the drawing]

Figure 1 is a diagram explaining the conventional phoneme recognition method, Figure 2
The figure is a diagram explaining a conventional similarity calculation method, and FIG. 3 is a complete explanation 19a of a method for creating a surrounding information standard pattern using the phoneme type (R) method in an embodiment of the present invention. -0°, FIG. 4 is a diagram illustrating a method for detecting feature parts and phoneme discrimination according to the same method according to the present invention, and FIG. 6 is a block diagram for fully realizing the same method according to the present invention. be. 1...Feature parameter extraction unit, 2...
Similarity calculation unit, 3...Standard pattern storage unit, 4
... Segmentation section, 6... Phoneme discrimination section. Name of agent: Patent attorney Toshio Nakao and 1 other person No. 1
Figure 2 Figure 3 Figure 4

Claims (1)

[Claims] (1) A standard pattern (hereinafter referred to as a phoneme standard pattern) created for each phoneme for a portion (hereinafter referred to as a feature part) that well expresses the characteristics of a phoneme in a phoneme interval, and identification. Prepare a standard pattern (hereinafter referred to as the surrounding information standard pattern) created for the surrounding information of the feature for the target phoneme group, perform all the segmentation of the input speech to find the feature part candidate section in the phoneme section, Pattern matching is performed by applying the phoneme standard pattern and surrounding information standard pattern to each point in the feature candidate section, and the similarity with each phoneme is determined by removing the influence of the surroundings of the feature. 1. A phoneme discrimination method, characterized in that phonemes are discriminated by comparing the degree of similarity within the feature section candidate sections over the entire candidate section. C2) As a distance measure in pattern matching, 2
The phoneme discrimination method according to claim 1, characterized in that a page statistical distance measure is used. (3) The phoneme discrimination method according to claim 2, wherein the statistical distance measure is Mahalanobis distance.