JPH042199B2 - - Google Patents

Description

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

Configuration of conventional examples and their problems In recent years, research and development on speech recognition for unspecified speakers and multiple words 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.
This method is suitable for recognition of a large number of words by any speaker, since the capacity of the word dictionary is 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. In particular, segmentation and recognition of consonants is a technically difficult problem.

Many studies have been conducted to clarify the characteristics of individual consonants or consonant groups, but there are many conventional examples of so-called automatic recognition, which identifies phonemes by segmenting consonants from a speech signal. do not have. A typical conventional method is to perform consonant recognition using the local peak of the spectrum as a feature parameter. magazine 34
(1978). However, this method does not have a sufficient consonant recognition rate. Here, as a conventional example, we will take up and explain the technology that the present applicant previously applied for.
List the problems.

Conventionally, the following three types of information are used for phoneme segmentation.

B Voiced/unvoiced/silent determination results The input audio is divided into frames, and each frame is determined as voiced, unvoiced, or silent, and is expressed as a time series. One frame is approximately 10msec. A consonant section is defined as a continuous portion of silence or unvoiced frames.

B Phoneme discrimination results using standard patterns of vowels, nasals, and unvoiced sounds 5 vowels, nasal sounds (/m/, /n/, hatsu sounds combined), voiceless (/s/, /h/ combined) Using seven standard patterns,
The input audio is compared with a standard pattern for each frame, and the standard pattern name with the maximum similarity is assigned to each frame and expressed as a time series. The consonant section is a section in which nasal frames or unvoiced frames persist.

C. Power depth Determine the audio power for each frame and express it as a time series. The consonant section is the part where the power dip occurs (power dip).

Phoneme segmentation is performed using the above three types of information as parameters. A specific example is shown in FIG.

FIG. 1 shows the movement of each parameter when ``Rakuda'' (/rakuda/) is uttered.
In the figure, a is the phoneme named manually, b is the voiced V, unvoiced u, silent Q judgment result, and c
shows the determination results for vowels (A, I, u, E, o), nasal sounds N, and silent sounds s for each frame. Further, d indicates the temporal movement of audio power. In each parameter, the section indicated by ←→ is determined to be a consonant section, and e indicates the determination result. Figure 1a was obtained by visual inspection, and it can be seen that the consonant part of a is detected with each parameter. Also,
If we combine the consonant intervals determined by each parameter, a
It almost matches the consonant interval of .

Next, phoneme discrimination is performed by comparing the detected consonant section with a consonant standard pattern frame by frame. Standard consonant patterns include: voiced consonants: /N/ (nasal), /B/ (voiced plosive /b/ /d/ /g/), /r/, /η/, /
h ₁ /(/a/, /o/, /u/ followed by /
h/), /h ₁ /(voiceless plosive /p/, /t/, /k/ followed by /a/, /o/, /u/) Voiceless consonant: /S/ (voiceless fricative /s/ ,/
c/), /h ₂ /(/h/ followed by a phoneme other than /a/, /o/, /u/), /k ₂ /(/a/, /
Voiceless plosives followed by phonemes other than o/, /u//
p/, /t/, /k/). For the consonant section, the voiced consonant standard pattern is applied to frames that are determined to be voiced in the voiced/voiceless/non-voiced determination results.
A standard unvoiced consonant pattern is applied to frames determined to be unvoiced, and the degree of similarity of each phoneme to the standard pattern is determined for each frame. Then, the degrees of similarity for each phoneme standard pattern are added up for all frames of the consonant section, and the phoneme name of the standard pattern with the largest sum is taken as the phoneme discrimination result of that consonant section. However, frames determined to be silent are not targeted. Furthermore, whether the standard pattern is determined as /k ₁ / or /k ₂ /, they are treated as the same and replaced with /k/. The same applies to /h/.

FIG. 1e shows an example of phoneme discrimination results, and it can be seen that relatively good results can be obtained.

The problem with the conventional example is that after determining an interval by segmentation, for the entire interval,
The point is that similarity calculation is performed for each frame. 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 are characterized by parameter movements throughout the phoneme interval.

Therefore, an effective method for distinguishing between consonants and semi-vowels is to extract 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 is not taken.

Purpose of the Invention The present invention solves the above-mentioned drawbacks of the prior art. First, phonemes are roughly classified into phoneme groups, and then the characteristic parts of each phoneme group are extracted, and the parameters of the characteristic parts are temporally determined. This method provides a means for identifying phonemes with high accuracy by performing matching with a standard phoneme pattern while taking into account the movement of the phoneme.

Composition of the Invention The present invention achieves the above object by segmenting input speech to determine phoneme intervals, and dividing the phoneme intervals by the magnitude of the high-frequency power dip caused by time fluctuations in the high-frequency components of the voice. The phonemes are roughly classified into four phoneme groups: voiced plosives, voiceless plosives, nasals, and fricatives using the magnitude of the high-frequency power dip and the magnitude of the low-frequency power dip in the phoneme interval. The size of the region power dip is used to automatically extract feature part (parts that are effective for phoneme discrimination) candidate sections, and the temporal The present invention provides a phoneme recognition method that calculates similarity by applying a standard pattern including a change process and a pattern of surrounding information of a feature, and simultaneously extracts an accurate feature and discriminates a phoneme. .

Description of examples The outline of this example is as follows.

B. Creation of standard phoneme patterns Phonemes are roughly classified into phoneme groups as follows, depending on the position of their characteristic parts. voiced plosives (/
p//t/, /k/, /c/), voiceless plosives (/b/, /d/, /g/), nasals (/
m/, /n/, /η/), voiceless fricative group (/
s/, /h/) However, the fluid sound (/r/) and nasal turbidity (/η/) are mixed into both the voiced plosive group and the nasal group, and the voiced fricative (/z/) is included in the voiced plosive group,
Mixed into the voiceless friction group.

A feature part is set for each phoneme group, and a phoneme standard pattern for each phoneme is created in advance for the feature part. The phoneme standard pattern is created using a large amount of data that is accurately labeled by visual inspection. In addition to the phoneme standard pattern, one type of standard pattern for surrounding information of the characteristic part is created for each phoneme group.

B. Phoneme discrimination The input speech is segmented to find phoneme intervals. Then, a part of the phoneme interval (for example, an end point) is set as a reference point. On the other hand, it is determined which phoneme group this phoneme section belongs to among the major classifications in A above. Next, the standard pattern belonging to the determined phoneme group is applied to the feature part in the phoneme interval to discriminate phonemes. By the way, it is generally difficult to automatically and accurately determine the features, so
Do as follows. That is, with reference to the above-mentioned reference points, a candidate section of the characteristic part is determined with some width, and a standard pattern is applied to the entire range of the candidate section to calculate the degree of similarity with each phoneme.
When calculating the degree of similarity with each phoneme, the degree of similarity with the standard pattern of the surrounding information of the phoneme group described in A above is removed from the degree of similarity between the phoneme standard pattern and the unknown input. By doing this,
Information on parts that do not correspond to the characteristic part (that is, parts corresponding to the periphery of the characteristic part) of the candidate section of the characteristic part can be removed, and phonemes can be discriminated by accurately capturing the characteristic part.

Hereinafter, one embodiment of the present invention will be described in detail with reference to the drawings, taking consonant recognition as an example.

FIG. 2 is a diagram for explaining a method of consonant recognition. In the figure, an input audio signal enters a filter analysis power calculation section 1 and an LPC analysis section 2. In the filter analysis and power calculation section 1, the phoneme signal is
Frequency analysis is performed using band filters for three channels, middle and high, and power (band power) is calculated for each channel for each frame (10 msec). The low-pass filter is 250-600Hz, the mid-pass filter is 600-1500Hz, and the high-pass filter is 1500-4000Hz.
It uses a certain amount of bandwidth. These band powers are mainly used to detect consonants and determine consonant intervals (perform consonant segmentation).

The LPC analysis unit 2 performs LPC analysis (linear predictive analysis) on the input audio signal for each frame.
The order of the analysis filter is approximately 15. The feature parameter extraction unit 3 calculates LPC cepstral coefficients from the results of the LPC analysis unit 2. This is a parameter that describes the outline of the frequency spectrum, and is an effective parameter for phoneme recognition for unspecified speakers. (Refer to Niyada et al., “Evaluation of feature parameters and distance measures for phoneme recognition of unspecified speakers,” Acoustical Society of Japan lecture, October 1982). vowel discrimination,
The voiced/unvoiced discriminator 5 distinguishes between vowels and nasal sounds (/
For each frame, the similarity of the feature parameters is calculated for the standard pattern of m/, /n/, a collection of nasal sounds), and the phoneme with the highest similarity is selected for that frame (targeting vowels and nasal sounds). Output as the determination result. Similarly, the degree of similarity to the voiced/unvoiced standard pattern is calculated, and the voiced/unvoiced discrimination result is output for each frame. The following statistical distance measure is effective as a distance measure for calculating similarity. (References above).

Input _{feature parameters} : _〓 = _{( X 1} , X _{2 , ...} Σ=6 ₁₁ 6 ₂₁ 〓 6 ₂₁ … … … … 6 _1,d 6 _d, dwhere d is the number of dimensions.

Then, the distance to phoneme i is defined as follows.

Distance based on Bayesian judgment: P _i =1/(2π) ^d/2・|Σ| ^1/2 exp{−1/2(〓−μ _i )
^T・Σ ^-1 _i・(〓−μ _i )} ...Formula 1 Mahalanobis distance: L _i = (〓−μ _i ) ^T・Σ ^-1・(〓−μ _i ) ...Formula 2 The subscript −1 is T represents the inverse matrix, and T represents the transpose.

In Equation 1, the phoneme that maximizes P _i and in Equation 2, the phoneme that minimizes L _i is the discrimination result for that frame. The same goes for voiced/unvoiced determination. Almost similar results can be obtained using Equations 1 and 2.

The consonant detection section 4 performs filter analysis and detects a power dip from the time-series waveform of the band power output from the power calculation section 1, and performs consonant segmentation. Furthermore, when the vowel discrimination and voiced/unvoiced discrimination section 5 includes consecutive frames determined to be nasal or unvoiced, these portions are also segmented as consonant sections.

The method of detecting the power dip will be explained with reference to FIG. FIG. 3a shows a time-series waveform of band power, and shows that concavities occur in consonant intervals. When a is differentiated (difference on a computer), it becomes b. In b, the interval l from the minimum value to the maximum value is a consonant interval. Further, the value d between the peak values of b is defined as the size of the power dip. O is the reference point.

The consonant major classification section 6 roughly classifies the consonants of the portions segmented as consonants by the consonant detection section 4 using the magnitude of the power dip. Generally, the spectrum of voiced consonants is concentrated in the low frequency range, so a large dip is likely to occur due to time series information of high frequency power. Furthermore, since the spectrum of unvoiced consonants is concentrated in the high range, a large dip is likely to occur due to the low range power. Since voiceless plosives rise from silence, there is a large dip in both low-frequency and high-frequency power. Nasal sounds do not have a large dip in either of them, but since they are voiced sounds, the dip is larger in high-frequency power.

FIG. 4 shows the position of each consonant group on the P _L -P _H plane, where the magnitude of the low frequency power dip is P _L and the magnitude of the high frequency power dip is P _H.
In FIG. 4, power dips in which both P _L and P _H are small and are not segmented as nasal or voiceless are removed from the consonant candidates as additional Vs. By focusing on the magnitude of the low-frequency power dip and high-frequency power dip in this way, consonants can be broadly classified into voiceless plosives, voiced plosives, voiceless fricatives, and nasals. However, /Z/ is included in both the voiced plosive group and the voiceless plosive group. This is because /Z/ is a phoneme that has the characteristics of both a buzzing part and a frictional part. Additionally, /γ/ and /η/ are mixed into both voiced plosives and nasals. This is because these phonemes are strongly influenced by the vowels before and after them, and the size of the power dip varies depending on the environment in which they appear.

After narrowing down the candidates through the major classification in this way, the next step is a consonant classification section 7 that performs detailed classification within each consonant group. The consonant subclassification unit 7 includes a voiceless plosive discrimination unit 8, a voiced plosive discrimination unit 9, a nasal discrimination unit 10, and a voiceless fricative discrimination unit 11, corresponding to each consonant group. In this part, the degree of similarity between the output of the feature parameter extraction unit 3 and the phoneme standard pattern is determined, and consonants are determined by comparing the degrees of similarity for each phoneme.

Voiceless plosives and voiced plosives are characterized by the transition from the plosive point to the following vowel. Therefore, in order to perform subclassification within a voiceless plosive group or a voiced plosive group, it is necessary to perform a similarity calculation that takes into account the temporal movement around the plosive point. Nasal sounds are characterized by the transition to a vowel, and it is necessary to calculate the degree of similarity by taking into account the temporal movement of this part. Ruon/
γ/ is characterized by the spectral change and duration of the entire interval. /Z/ is characterized by having a bub part and a friction part following it.

As described above, although there are differences in the characteristic parts depending on each consonant group, it is common that the temporal movement based on the characteristic point is important information. It is not easy to automatically extract feature parts accurately. However, since it is known empirically that the characteristic part of each phoneme group is near the rise of the power dip,
As shown in Figure 2, using the rising frame of the power dip as a reference point, the degree of similarity is calculated over several frames before and after it, and the value of the frame with the maximum degree of similarity is taken as the degree of similarity for that phoneme. This part will be described later.

Next, regarding calculation of similarity, Formula 1 or Formula 2
is used to calculate similarity considering temporal movement. That is, instead of the feature parameters of a single frame, the feature parameters of a plurality of frames (here, 1 frame) are used as data used for similarity calculation. In equation 1 or _equation 2 _〓 = ₍ X ¹ ₁ , X ¹ ₂ ,... ^{X 1 d} _, X ^{2 1 ,}
₁ , X ^l ₂ , ...X ^l _d ) = (μ ¹ ₁ , μ ¹ ₂ ...μ ¹ ₂ ...μ ¹ _d μ ² ₁ , μ ² ₂ ...
d×l-dimensional data is used, such as μ ² _d … μ ^l ₁ , μ ^l ₂ … μ ^l _d ). Similarly, the covariance matrix is assumed to have d×l dimensions. (I won't write it down as it would be complicated). By using multiple frames of data in this way, it is possible to simultaneously compare the spectral characteristics of the parameters and the characteristics of their temporal fluctuations with the phoneme standard pattern.

Next, we will explain how to create a standard pattern. The standard pattern is created using a lot of data that is accurately extracted from the audio by visual inspection.

The phoneme standard pattern is obtained by cutting out l frames of data corresponding to characteristic parts from many data of the same phoneme, and obtaining a d×l dimension characteristic spectrum.
It is created for each phoneme by finding the average value and covariance matrix of a lot of data.

One standard pattern of surrounding information is created for each phoneme group. This is because surrounding information is common to each phoneme within a phoneme group.
For example, voiced plosives (/b/, /d/, /
g/), there are always several frames of buzz before the feature (rupture), and after the burst, it connects to a vowel. The standard pattern of surrounding information is thus a standard pattern of universal surrounding information for that phoneme group. Fig. 5 shows the method of making it. In the vicinity of the characteristic part (hatched area in the figure), a section that is sufficiently long in time compared to the characteristic part is set as the surrounding information section L. For this section,
As shown in the figure, the feature parameters (d×l dimension) of l frames are extracted over the entire interval while being shifted one frame at a time. Such a procedure is applied to a lot of data belonging to the same phoneme group to obtain an average value vector and a covariance matrix, and this is used as a standard pattern of surrounding information. In this way, although the standard pattern of surrounding information includes data on the characteristic part, the data in the vicinity of the characteristic part has a much greater weight than that.

Next, a specific method will be described for subclassifying data that has been roughly classified into phoneme groups by the method shown in FIG. 4, using the standard pattern created by the above method.
In the following explanation, the distance measure of Equation 2 will be used for simplicity, and a case will be discussed in which one phoneme group is composed of two phonemes (phoneme 1, phoneme 2). The idea remains the same even if the number of phonemes increases.

As mentioned above, for the feature section, a rough candidate section is determined based on the rising frame of the power dip. This interval is defined as t ₁ to _{t 2} in terms of time.
The unknown input vector (data to be subdivided) at time t is now t (t=t ₁ ~ _{t 2} ) The standard pattern of phoneme 1 (average value) is the standard pattern of ₁ phoneme 2 (average value) ₂ Let the standard pattern (average value) of surrounding information be 〓 ₂ , 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.

If the similarity (distance) to unknown input phoneme 1 at time t is L ₁ , t, then L ₁ , t = (〓t-〓 ₁ ) ^T・Σ ^-1・(〓t-〓 ₁ )
−(〓t−〓 _e ) ^T・Σ ⁻¹・(〓t−〓 _e ) Equation 3 Similarly, if the distance to phoneme 2 is L ₂ , t, then L ₂ , t=(〓t−〓 ₂ ) ^T・Σ ^-1・(〓t−〓 ₂ )
−(〓t−e) ^T・Σ ⁻¹・(〓t−〓 _e ) Formula 4. What these expressions mean is that the time t
The similarity between the unknown input and the phoneme standard pattern minus the similarity with respect to the surrounding information is set as the new similarity with the phoneme. and formula 3
The calculation of Equation 4 is performed for the period from t ₁ to _{t 2} , and the phoneme that is the smallest among L ₁ , t, L ₂ , and t during this period is determined as the recognized phoneme.

In reality, Equations 3 and 4 can be expanded into simple equations as follows. (Derivation omitted).

L ₁ , t=〓 ₁・〓t−〓 ₁ formula 3′ L ₂・t=〓 ₂・〓t−〓 ₂ formula 4′ 〓 ₁ , 〓 ₂ , 〓 ₁ , 〓 ₂ are standards that include surrounding information It's a pattern.

The meaning of the above method will be conceptually explained with reference to FIG.

Consider the case where consonant discrimination is performed in a situation where the phoneme section is shown in FIG. 6a. For the true characteristic part (hatched part) of this consonant, the characteristic part candidate section T is time
It is assumed that the values are obtained as t ₁ to t ₂ . b is phoneme 1 (solid line), phoneme 2 (diagonal line), calculated by equation 2.
This figure shows temporal fluctuations in the degree of similarity. A, B, and C indicate positions where the degree of similarity is minimal. 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 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. 6c 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. FIG. 6 d is the distance to the phoneme standard pattern containing 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.

As described above, by using the method of this embodiment, it is possible to automatically extract true features accurately and discriminate phonemes from the rough feature candidate sections determined by the segmentation parameters. can.

In addition, although the Mahalanobis distance based on Formula 2 was explained above, a similar method can be used for other distances. For example, Equation 1 can be treated in the same way as Equation 2 by taking logarithms. (In this case, likelihood is calculated instead of distance)
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.

In this way, the method narrows down the number of candidates through major classification, and uses automatically extracted feature parts as a basis for subclassification, and uses a statistical distance scale to determine phonemes by considering temporal movement. It is a rational recognition method that utilizes the phonetic properties of consonants and semi-vowels in particular.

For convenience of explanation, Fig. 2 describes a format in which segmentation is first performed, a reference point is detected, and then similarity is calculated, but when it is actually implemented into a device, it is necessary to perform segmentation while shifting one frame at a time. , similarity is calculated centering on all frames, segmentation is performed in parallel to find a reference point, and phonemes are classified by referring to the similarity with respect to the reference point, but the essence is There is no difference.

In this example, all middle consonants (/p/, /
t/, /k/, /c/, /b/, /d/, /
η/, /m/, /n/, /γ/, /z/, /
s/, /h/), an average recognition rate of about 76.1% was obtained. The data uses words generated by a total of 20 men and women, and is sufficiently reliable.
When evaluating the conventional method under similar conditions, the consonant (/
The average recognition rate is about 72.5% for γ/, /η/, /h/, /s/, /c/ and consonant groups (voiceless plosive group, voiced plosive group, nasal group). Considering that the conventional method does not perform subdivision within some consonant groups, it can be seen that the effects of the embodiments of the present invention are significant.

In addition, in the case of subclassification of consonants, when using standard patterns that do not include surrounding information, 72.7% of voiced plosives (/b/, /d/ and /η/) in words, and 72.7% of nasals (/m/) /, /n/ and /η/) 64.1%
It was hot. When using a standard pattern that includes surrounding information, this increases to 74.7% and 75.4%, respectively. In particular, a remarkable effect appears for the nasal group. This is because the power dip of nasal sounds is unclear, so the reference point cannot be detected accurately.

Effects of the Invention In summary, the present invention performs segmentation of input speech to determine phoneme intervals, and divides the phoneme intervals into the magnitude of the high-frequency power dip and the magnitude of the low-frequency power dip caused by time fluctuations in the high-frequency components of the voice. The phonemes are classified into four phoneme groups: voiced plosives, voiceless plosives, nasals, and fricatives.
Next, a feature part (part effective for phoneme discrimination) candidate section is automatically extracted using the magnitude of the high-frequency power dip and the magnitude of the low-frequency power dip in the phoneme section, and the feature part candidate section is Calculate the degree of similarity by applying a standard pattern that includes the temporal change process of individual phonemes belonging to a group of broadly classified phonemes and the surrounding information pattern of the feature to extract accurate features and discriminate between phonemes. The present invention provides a phoneme recognition method that is characterized by: (a) Automatic segmentation of speech and the ability to recognize phonemes with high accuracy.

(b) It is possible to automatically and accurately extract parts (feature parts) that are effective for phoneme discrimination and perform matching.

C. Subclassification within voiced plosives, voiceless plosives, and nasals, which have been difficult to distinguish in the past, can be performed in combination with automatic segmentation.

By using a relatively simple parameter called the power dip, it is possible to efficiently classify the segmentation consonants and detect reference points for similarity calculation.

It has the following advantages.

[Brief explanation of drawings]

FIG. 1 is a diagram explaining a conventional phoneme segmentation method, FIG. 2 is a block diagram explaining a phoneme recognition method according to an embodiment of the present invention, and FIG. 3 is a diagram showing the power dip and its magnitude of the same embodiment. FIG. 4 is a diagram conceptually explaining the method of consonant classification according to the same embodiment, and FIG. 5 is a diagram explaining the method of creating the surrounding information standard pattern according to the same embodiment. 6 are diagrams for explaining a method for detecting characteristic parts and discriminating phonemes in the same embodiment. 1...Filter analysis/power calculation section, 2...LPC
Analysis unit, 3... Feature parameter extraction unit, 4... Consonant detection unit, 5... Vowel discrimination/voiced/unvoiced discrimination unit, 6... Consonant major classification unit, 7... Consonant subclassification unit, 8... Voiceless plosive discrimination unit, 9 ...voiced plosive discriminator, 10... nasal discriminator, 11... voiceless fricative discriminator.

Claims (1)

[Scope of Claims] 1. Segmentation of input speech is performed to determine phoneme intervals, and the phoneme intervals are determined by combining the magnitude of the high-frequency power dip and the magnitude of the low-frequency power dip caused by time fluctuations in the high-frequency components of the voice. The phonemes are roughly classified into four phoneme groups: voiced plosives, voiceless plosives, nasals, and fricatives, and then using the magnitude of the high-frequency power dip and the magnitude of the low-frequency power dip in the phoneme interval. A feature section (a part effective for phoneme discrimination) candidate sections is automatically extracted using the method, and a standard pattern including the temporal change process of each phoneme belonging to the broadly classified phoneme group is extracted for the feature section candidate section. A phoneme recognition method characterized by calculating similarity by applying surrounding information patterns of feature parts, extracting accurate feature parts, and simultaneously discriminating phonemes. 2. The phoneme recognition method according to claim 1, wherein similarity calculation is performed using a statistical distance measure as a distance measure in matching with a standard pattern that includes a temporal change process of phonemes.