JPH0316039B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION Field of Industrial Application The present invention relates to a phoneme recognition method in a phoneme recognition method characterized by performing phoneme recognition.

Configuration 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, which are characterized by phoneme recognition, are less susceptible to speaker-specific variations such as differences in accents, and convert speech signals into phoneme sequences, which contain a small amount of information and are compatible with linguistics. The capacity of the word dictionary may be small;
This method is suitable for recognizing a wide variety of speakers and a large number of words, as 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.

Phonetic research on the characteristics of individual consonants is
It has been done many times before. However, there are not many conventional examples of so-called automatic recognition, a method of segmenting a speech signal to detect consonant intervals and recognizing phonemes. Here, we will take up as a conventional example a technique previously applied for by a group including the present applicant, and discuss its 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 unvoiced sections as consonant sections. As shown in Figure 1a, the audio
It is divided into frames of about 10 msec, and each frame is judged to be voiced or unvoiced. Then, a section in which unvoiced frames are continuous is defined as a consonant section. This method is effective for detecting voiceless consonants.

(b) Each frame of the speech section is divided into five vowels and a nasal sound (/
The phoneme with the highest degree of similarity is selected as the recognition result for that frame. Then, a section in which consecutive frames recognized as nasal sounds are defined as a consonant section (FIG. 1b). 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 from the temporal movement of voice power, and define the dip section as a consonant section (Fig. 1c). 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, consonants are recognized in the consonant interval obtained in this manner as follows. Second
This will be explained using figures. The degree of similarity between each consonant and the standard pattern is calculated for each frame of all frames in the consonant section. In the figure, the degree of similarity to phoneme A is expressed as la, the degree of similarity to phoneme B is expressed as lb, and the degree of similarity to phoneme C is expressed as lc. The degree of similarity is calculated using (Equation 1).

l _j =−(x−1μ _j ) ^T・Σ ⁻¹ _j・(x−1μ _j )−K _j
...(Equation 1) However, j is a phoneme name (a, b, c...), and K _j is a constant depending on phoneme j. Also, x is the input feature parameter (LPC cepstral coefficient) vector,
1μ is the mean value vector (standard pattern), and Σ is the covariance matrix (standard pattern).

Then, phonemes A, B, C... for all frames.
Let the sum of similarities be L _a , L _b , L _c , respectively, then L _a = _K 〓 ⁱ⁼¹ l _a,k , L _b = _K 〓 ⁱ⁼¹ l _b,k , L _c = _K 〓 ⁱ⁼¹ l _c,k ……. In this way, the phoneme with the largest similarity sum 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 temporally static, and all intervals are treated equally.

However, 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 phoneme discrimination are concentrated near the plosive, and in nasal sounds, the features for phoneme discrimination are located in the transition to the following vowel.
Flowing 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 distinguishing 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 the characteristic parts of phonemes and match them with the standard phoneme pattern, including the temporal movement of parameters in the characteristic parts. By doing so, the purpose is to provide a means for identifying phonemes with high accuracy.

Structure of the Invention The present invention achieves the above object, and consists of 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 "surrounding information standard pattern") created for the surrounding information of the characteristic part regarding the phoneme group to be identified, and segmentation of the input speech is performed to find the information in the phoneme section. Find feature candidate sections, perform pattern matching by applying the phoneme standard pattern and surrounding information standard pattern to each point in the feature candidate section, and compare the degree of similarity with each phoneme to the features surrounding the feature section. This invention provides a phoneme discrimination method characterized in that phonemes are discriminated by determining the entire feature part candidate section in a form in which the influence is removed and comparing the degree of similarity within the feature part candidate section.

DESCRIPTION OF EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings.

In this example, the characteristics of each phoneme are accurately detected visually, a standard phoneme pattern is created in advance using a large amount of data, and all phonemes (phoneme groups) are compared for similarity at once. A standard pattern of surrounding information of the feature part is also created for phonemes belonging to ), and phoneme discrimination is performed by first setting the candidate interval of the feature part a little wider using the segmentation parameter, and then For the entire candidate section, set one standard pattern for each phoneme and one standard pattern for surrounding information.
By applying this technique while shifting each frame and calculating the degree of similarity, features can be accurately detected from candidate sections and phonemes can be discriminated.

First, a method for creating a standard pattern will be explained. The phoneme standard pattern is created by extracting feature parameters from visually extracted feature parts for each phoneme, and finding the average value and covariance matrix of a large amount of data. The feature parameters are LPC cepstral coefficients obtained by linear predictive analysis (LPC analysis).

Now, if the number of feature parameters per frame is d and the number of frames of the feature part is j, then the parameter series x of the feature part is x = (x ⁽¹⁾ ₁ , x ⁽¹⁾ ₂ ...x ⁽¹⁾ _d , x ⁽²⁾ ₁ , x ⁽²⁾ ₂ , x ⁽²⁾ _d
, ... x ^(j) ₁ , x ^(j) ₂ ... x ^(j) _d ) ... (Formula 2). x ^(k) _j is the i-th LPC cepstral coefficient in the k-th frame of the feature. Parameter sequences are extracted from a large amount of data, and the average value vector 1μ of each element and the covariance matrix Σ between elements are determined and used as a standard pattern.

1μ=(μ ⁽¹⁾ ₁ , μ ⁽¹⁾ ₂ , …μ ⁽¹⁾ _d , μ ⁽²⁾ ₁ , μ ⁽²⁾ ₂
,... μ ^(j) ₁ , μ ^(j) ₂ ...μ ^(j) _d ) (Formula 3) where μ ^(k) ₁ = _M 〓 ^l=1 xμ ^(k) _i l/M (M is the number of data ) Since the covariance matrix is complex, it is not described here. As described above, the method of this embodiment is characterized in that a standard pattern is created using feature parameters of a plurality of frames, that is, by taking into account 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 plosives (/b/, /
d/, /q/), one for the nasal group (/
The ratio is one for every m/, /n/, /η/).

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 of sufficient length before and after the characteristic part (the shaded section in the figure). The length of the interval is set for each phoneme group. For the surrounding information section L, a d×l dimensional parameter series is obtained for the entire section by dividing it into l frames and shifting it by one frame as shown in the figure. Apply this operation to all data belonging to the phoneme group,
A standard pattern for surrounding information is obtained 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, determine the consonant interval using the segmentation parameter,
Using the rear end of the consonant section as a reference point, the candidate section of the characteristic part is set to be somewhat wider. Although any segmentation parameter may be used, in this embodiment, the power dips of the high and low frequencies are mainly used, and the rising portion of the dip is used as the reference point. When the power dip is difficult to detect, the voiced/unvoiced judgment result and nasality are used together.

Next, a method of detecting a feature part from a feature part candidate section and determining a phoneme will be explained. 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.

Let the characteristic part candidate section be t ₁ to _{t 2} . Now at time t
Let x _t be the unknown input vector (data to be determined) at (t ₁ ≦t≦t ₂ ). x _t has the same format as (Equation 2). Then, the standard pattern for phoneme 1 (average value) is 1μ ₁ , the standard pattern for phoneme 2 (average value) is 1μ ₂ , and the standard pattern for surrounding information (average value)
Let be 1 μ _e , and let Σ be the covariance matrix common to all of phoneme 1, phoneme 2, and surrounding information. Σ is created by averaging each covariance matrix.

Let L _1,t be the similarity (distance) between the unknown input and phoneme 1 at time t, then L _1,t = (x _t −1μ ₁ ) ^T・Σ ⁻¹・(x _t −1μ ₁ ) −(x _t −1μ _e ) ^T・Σ ⁻¹・(x _t −1μ _e ) …(Formula 4) Similarly, if the distance to phoneme 2 is L _{2 .t} , then L ₂ , _t = (x _t −1μ ₂ ) ^T・Σ ^-1・(x _t −μC ₂ ) −(x _t − μ _e )・Σ ⁻¹・(x _t − μ _e )...(Equation 5). 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. Calculate (Equation 4) and (Equation 5) from t ₁ to _{t 2}
The calculation was conducted for the period of , and out of L ₁ , _t and L ₂ , _t ,
The phoneme that is the smallest during this period is the recognized phoneme.

In reality, (Formula 4) and (Formula 5) can be expanded into a simple formula as follows (the derivation is omitted as it is not essential).

L ₁ , _t = α ₁・x _t − |B ₁ … (Equation 4)′ L ₂ , _t = α ₂・x _t − |B ₂ … (Equation 5)′ α ₁ , α ₂ , |B ₁ , | Set B ₂ as a new standard pattern that includes surrounding information.

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

Consider the case where consonant discrimination is performed in the situation shown in FIG. 4a. Assume that the characteristic part candidate section T is calculated as _t1 to _t2 for the true characteristic part (shaded part) of this consonant. b shows temporal fluctuations in the distances to phoneme 1 and phoneme 2 using solid lines and broken lines, respectively. A, B, and C indicate positions where the distance is minimum. In the true feature (point B), 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 using the segmentation parameters, phoneme 2 is at its minimum at point A, so if this continues, it will be misclassified as phoneme 2. FIG. 4c shows the distance between the unknown input surrounding information and the standard pattern, and the value increases near the true feature. This is because the standard pattern is created mainly using peripheral information. FIG. 4 d is the distance from the standard phoneme pattern that includes surrounding information, and is equivalent to b minus c. At point d, the value at point B is smaller than at point A, and this consonant is correctly determined as phoneme 1.

In this way, 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 phonemes.

Note that although the explanation was given above using the Mahalanobis distance, the concept is the same for other distances. For example, when using (Equation 1), the likelihood may be used instead of the distance, and the local maximum value may be used instead of the local minimum value. Further, although the above explanation has been made using consonants, a similar method can also be applied to phonemes that change over time, such as semi-vowels.

FIG. 5 is a block diagram for implementing the method of the invention. Reference numeral 1 denotes a feature parameter extraction unit, which analyzes input speech and extracts feature parameters.

Here, LPC analysis is performed to obtain LPC cepstral coefficients and send them to the similarity calculation unit 2. Also,
The feature parameter extraction unit 1 calculates the high-frequency power and low-frequency power of the input audio, and then extracts the high-frequency power and low-frequency power of the input voice.
Send to. The similarity calculation unit 2 calculates the distance between the input parameter and each standard pattern stored in the standard pattern storage unit 3 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. The segmentation unit 4 determines phoneme intervals based on the similarity information sent from the similarity calculation unit 2 and the power information sent from the feature parameter extraction unit 1. The phoneme discrimination unit 5 extracts a feature part based on the similarity between the phoneme interval and each phoneme, and the similarity of surrounding information,
The similarity of phonemes in the feature parts is compared, phonemes are discriminated, and the results are output.

In this example, we can use the average consonant to
A recognition rate of 76.1% was obtained. When similar evaluation was performed using the conventional method, it was 72.5%. 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 212 words uttered by a total of 20 men and women, and the reliability of the above results is sufficient.

In addition, to investigate the effect of using standard patterns of surrounding information, we conducted experiments with voiced plosives and nasals. As a result, the recognition rate for vocal plosives is low when the standard pattern of surrounding information is not used.
72.7% and 64.1% for nasal sounds, while using the standard pattern of surrounding information, 74.7% and 64.1% for nasal sounds, respectively.
This improved to 75.2%. In particular, a remarkable effect appears on nasal sounds. This is because the power dip of the nasal sound is unclear, so the phoneme interval cannot be extracted accurately. When the standard pattern of surrounding information was introduced, even when the phoneme section was unclear, the characteristic part could be detected accurately and the recognition rate improved, thereby verifying the effect of this example.

Effects of the Invention In summary, the present invention provides a standard pattern created for each phoneme for a portion that well expresses the characteristics of a phoneme (hereinafter referred to as a feature) in a phoneme interval, and a standard pattern around the characteristic portion for a group of phonemes to be identified. A standard pattern created for the information is prepared, segmentation of the input speech is performed to obtain feature part candidate sections in the phoneme section, and the phoneme standard pattern and the above phoneme standard pattern are used for each point in the feature part candidate section. Pattern matching is performed by applying the surrounding information standard pattern, and the degree of similarity with each phoneme is determined for the entire feature candidate section in a form that removes the influence of the surroundings of the feature, and the similarity within the feature candidate section is calculated. This provides a phoneme discrimination method characterized by discriminating phonemes through comparison. (a) It is possible to perform automatic segmentation of speech and discriminate phonemes with high accuracy.

(b) By using the phoneme standard pattern and the surrounding information standard pattern, it is possible to automatically and accurately extract parts (characteristic parts) that are effective for phoneme discrimination and perform matching.

It has the following advantages.

[Brief explanation of the drawing]

FIG. 1 is a diagram explaining a conventional phoneme recognition method, FIG. 2 is a diagram explaining a conventional similarity calculation method, and FIG. 3 is a diagram of a surrounding information standard pattern of a phoneme discrimination method in an embodiment of the present invention. FIG. 4 is a diagram explaining the method for detecting characteristic parts and discriminating phonemes according to the same method according to the present invention. FIG. 5 is a block diagram for realizing the method according to the present invention. It is a diagram. 1... Feature parameter extraction unit, 2... Similarity calculation unit, 3... Standard pattern storage unit, 4... Segmentation unit, 5... Phoneme discrimination unit.

Claims (1)

[Scope of Claims] 1. A standard pattern (hereinafter referred to as a phoneme standard pattern) created for each phoneme for a part of the phoneme interval that well expresses the characteristics of the phoneme (hereinafter referred to as a feature part), and a recognition target. A standard pattern (hereinafter referred to as the surrounding information standard pattern) created for the surrounding information of the feature for the phoneme group to be used is prepared, and the input speech is segmented to find the feature 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 degree of 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 determining the entire feature part candidate section and comparing the degree of similarity within the feature part candidate section. 2. The phoneme discrimination method according to claim 1, characterized in that a statistical distance measure is used as a distance measure in pattern matching. 3. The phoneme discrimination method according to claim 2, wherein the statistical distance measure is Mahalanobis distance.