JPH01260499A - Consonant recognizing method - Google Patents

Abstract

PURPOSE:To improve the accuracy of a large classification of a consonant and the recognition rate by applying the magnitude of a dip of low band power and high band power of an input voice spectrum to a discrimination chart of the succeeding separate vowel and classifying it into four consonant groups and several intermediate areas. CONSTITUTION:A power dip is detected from low band power and high band power by a power dip detecting part 10, and magnitude of the dip is derived by a power dip magnitude extracting part 11. Subsequently, with respect to each phoneme group, a standard pattern consisting of a variance, covariance and an average value is generated by using the magnitude of the power dip of the low band and the high band as a parameter, and a discrimination chart for executing in advance a discrimination to the phoneme group of the highest similarity against each input is generated. Next, when the consonant is divided roughly into four phoneme groups by using this discrimination chart, an erroneous classification is generated in the vicinity of a boundary of the discrimination chart, and the recognition cannot be executed correctly. Therefore, an intermediate area is set in the vicinity of vicinity of a boundary of the phoneme group, the consonant which has been classified into this intermediate area executes matching with both standard patterns of the adjacent phoneme groups and the consonant is recognized. In such a way, the recognition rate can be improved.

Description

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

BACKGROUND OF THE INVENTION Recently, consonant recognition methods have been widely used in the field of speech recognition. This consonant recognition method uses, for example, t, ``
"Word speech recognition system using the outline form of the speech spectrum and its dynamic characteristics" (Journal of the Acoustical Society of Japan, Vol. 34, No. 3, 197B)
) is known.

Hereinafter, a conventional consonant recognition method will be explained with reference to FIGS. 13 and 14.

First, the conventional word recognition device divides the input speech into phoneme units and recognizes them as a combination of phonemes (called phoneme recognition). A functional block diagram is shown in FIG. First, the audio analysis unit 1 analyzes the voices of multiple speakers using a filter bank for each frame (one frame is 10rns·C), and the feature extraction unit 2 analyzes the voices of multiple speakers using a filter bank based on the obtained spectrum information. Find the feature parameters by The standard pattern registration unit 3 creates standard patterns for each of the five vowel and consonant phoneme groups from these feature parameters.
Register in . When actually performing recognition,
A segmentation unit 4 performs segmentation of consonants using the feature parameters determined by the feature extraction unit 2.

Based on this result, the phoneme discrimination section 5 determines the phoneme by comparing it with the standard pattern of the standard pattern training section 3. Finally, the time series of phonemes created as a result is sent to the word recognition unit 6, and the word corresponding to the item with the highest degree of similarity to the word dictionary 7 similarly expressed in the time series of phonemes is output as a recognition result.

Here, for segmentation, when the shape of the temporal change 8 of the overall power is concave as shown in Fig. 14 (this is called a dip), the frame in which the power shows the minimum value is set as 01, and the frame where the power shows the minimum value is set as 01. A frame in which the rate of change in power over time 9 (this is called a power difference value) has negative and positive maximum values in the previous and subsequent frames is designated as n2. Let it be n3. Also, if the difference value at a certain frame n is WD(n), wo(n), wo(n 3 )
When satisfies the condition of WD (n 2 )≦θW D (n 3 )≦θW, the section from n2 to n3 is defined as a consonant section. Here, Ow is a threshold value for preventing addition of consonants.

Next, characteristic parameters indicating phoneme characteristics for this consonant section are determined for each frame, and consonant classification is performed for each frame by comparing it with a standard pattern for each phoneme for which a pattern has been prepared. This result is applied to the consonant classification tree to classify consonants into those that match the conditions.

Problems to be Solved by the Invention However, in the above configuration, first segmentation of mid-word consonants is performed using power dip, and then consonant classification is performed for each frame.

Finally, the results of the consonant classification for each frame are applied to the consonant classification tree, and the algorithm is extremely complex and time-consuming in order to classify the consonants into those that meet the conditions. Furthermore, since only the overall power is used as power information, there is also the problem that the accuracy of consonant segmentation is poor.

The present invention is intended to solve the above-mentioned problems of the prior art, and it is an object of the present invention to greatly simplify and accurately classify and recognize consonants in input speech.

Means for Solving the Problems The present invention applies the magnitude of the dip between the low-frequency power and the high-frequency power of the input speech spectrum to a discriminant diagram for each subsequent vowel.
The above objective is achieved by classifying consonants into two consonant groups and several intermediate regions.

Effect of the Invention With the above configuration, the present invention applies the size of the dip of low-frequency power and high-frequency power to the discriminant diagram for each subsequent vowel, thereby making it possible to easily and accurately classify consonants and improve the recognition rate. can be done.

Embodiments Hereinafter, embodiments of the present invention will be described with reference to the drawings.

In this example, the phonemes /p/, /l/, /ni/, /c/.

/b/, /d/, /m/, /n/, /s/, /h/ as silent plosives (/p/, /l/, /ni/, /c/) and voiced plosives (/p/, /l/, /ni/, /c/). /b/.

/d/'t, nasals (/rn/, /n/), voiceless fricatives (/, /.

An example will be described in which consonants are roughly classified into four phoneme groups (/h/) and the vicinity of the boundary between each phoneme group is regarded as an intermediate region.

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 is a functional block diagram embodying a consonant recognition method in an embodiment of the present invention. In FIG. 1, 10 is a power dip detection section, and 11 is a power dip magnitude extraction section, which obtains the magnitude of the dip detected by the power dip detection section 10. Reference numeral 12 denotes a vowel recognition unit for each frame, which performs vowel recognition for each frame from the end of the dip detected by the power dip detection unit 10. Reference numeral 13 denotes a subsequent vowel recognition unit, which recognizes subsequent vowels from the vowel recognition results for each frame obtained by the vowel recognition unit 12 for each frame. 14 is a discriminant diagram selection unit, and a subsequent vowel recognition unit 13
The discriminant diagram for the subsequent vowel recognized in is selected from the discriminant diagram storage unit 15 for each subsequent vowel.

Reference numeral 16 denotes a major classification determining section, which uses the discriminant diagram selected by the discriminant diagram selecting section 14 to determine the major classification of consonants. The standard pattern selection section 17 takes out a necessary standard pattern from the standard pattern storage section 18 based on the results of the major classification by the major classification determination section 16, and matches it with the standard pattern in the consonant recognition section 19 to perform consonant recognition.

The operation of the above configuration will be explained below.

In the present invention, low frequency power and high frequency power are used as characteristic parameters. Voiced consonants tend to have a power dip in high-frequency power, and voiceless consonants tend to have a power dip in low-frequency power.

Therefore, by using both low and high frequency power, it becomes possible to respond to all consonants. Furthermore, since the magnitude of the power dip is affected by the following vowel, accuracy will be improved if a discriminant diagram is created for each subsequent vowel.

Next, to explain specifically, first, a power dip is detected from the low frequency power and the high frequency power, and the size of the dip is determined. The second method for determining the magnitude of this power dip is
This will be explained using Figures (a) and (b). In the figure, n is the frame in which the temporal change rate 21 of the high frequency power has a positive maximum value.
Let n2 be the frame in which the temporal change rate 23 of l% low-frequency power reaches a positive maximum value. The magnitude of the rate of change in each frame is WD (n 1), WD (n 2
). The magnitude PL of the low frequency power dip and the magnitude PH of the high frequency power dip are defined as PH=WD (nl) PL=WD(n2).

Using this dip size, voiceless plosives (/p/,
/l/, /ni/, /c/), voiced plosive (/b/.

/d/), nasal C/m/, /, />, voiceless fricative C/m
/.

/h/) When examining the distribution of PL and PH, Figure 3~
It will look like Figure 6. In the figure, the horizontal axis is PL and the vertical axis is PH.
The numbers in the figure represent the number of phonemes that appear. As is clear from the figure, phonemes that exhibit plosiveness have large PL and PH, particularly voiceless plosives have large PL, and voiced plosives have large PH. In addition, phonemes that do not show plosiveness are PL, P
H is also small, but it can be divided into voiced or voiceless consonants as shown in Figures 5 and 6.

Therefore, by using the magnitude of the power dip in the low and high frequencies, consonants can be roughly classified. For each phoneme group, a standard pattern consisting of variance, covariance, and average value is created using PL - PH as parameters, and a discriminant diagram is created in advance to discriminate the phoneme group with the highest degree of similarity for each input. If consonants are roughly classified into four phoneme groups using this discriminant diagram, misclassification will occur near the boundaries of the discriminant diagram. In this example (/p/, /i/, /ni/,
/c/) is about 7% more than (/b/, /d/), (/b/, /
4/) is incorrectly classified as (,'p/, /l/, /ni/, /c/) by about 8.3%. If this continues, if the major classification is incorrect in the discriminant diagram, the incorrect phoneme group will be matched with the standard pattern, making it impossible to recognize it correctly.

Therefore, an intermediate region is set near the boundary of phoneme groups, and consonants classified into this intermediate region are matched with standard patterns of both adjacent phoneme groups to recognize consonants, thereby improving the recognition rate.

FIG. 7 shows an example of a discriminant diagram. In the figure, the horizontal axis represents the low frequency range, and the vertical axis represents the magnitude of the high frequency power dip.

The four areas divided by solid lines are the discrimination boundaries in the case where there is no intermediate reduction. The area I-4 surrounded by the dotted line is the intermediate area. When the magnitude of the power dip in the consonant section falls within this intermediate range, matching is performed with the phoneme standard pattern of the adjacent phoneme group. The intermediate region I is (/p/, /l
/, /ni/, /c/, /b/.

/d/), 1 is (/p/, /l/, /ni/, /c
/ , /s/.

/h/), It is (/b/, /d/, /, /, /h
/), IV is (/b/, /c+/, /m/, /n/)
, V is (/m/, /, /.

/3/, /h/) is matched with the phoneme standard pattern. For the four phoneme groups lζ other than the intermediate region, matching is performed with the standard pattern of each phoneme group.

Furthermore, when we examine the magnitude of the power dip for each subsequent vowel, we find that the magnitude of the dip is different even for the same consonant. Therefore, in order to improve consonant segmentation accuracy, subsequent vowel information is used to roughly classify consonants by subsequent vowel. As an example, FIGS. 8 to 12 show temporal change patterns of power for each subsequent vowel in the case of the phoneme /r/. Figure 8 is /, a/, Figure 9 is /rl/, Figure 10 is /, u/, Figure 11 is /, 6/, Figure 12 is /'0
The case of / is shown.

In the figure, the horizontal axis is time, the vertical axis is the magnitude of power, the solid line is the change in power over time, and the dotted line is the rate of change in power over time.
PL indicates the movement of low frequency power, and PH indicates the movement of high frequency power. Also, starting from Figure 3, the following vowels are /a/, /I/, /, /,
/, s/, 10/. Looking at the dip size by following vowel, /U/.

10/ (Figures 10 and 12), the magnitude of the high-frequency power dip is greater than that of other subsequent vowels (/a/, /I/).

/e/). This is because the power varies slightly depending on the vowel, resulting in a difference in the speed at which the power changes over time from a consonant to a vowel. ／``／
Similarly, when we examine the temporal changes in power for other consonants, we find that the magnitude of the power dip differs depending on the following vowel.

Therefore, accuracy can be improved if consonants are roughly classified by following vowel using the magnitude of the power dip in the low and high frequencies. The subsequent vowels are recognized frame by frame by calculating the similarity between the input voice data and a standard pattern of five vowels created in advance from a large amount of data. Figure 2 (
0) shows the first and second vowel recognition results for each frame. In order to easily determine the subsequent vowel, for example, the vowel recognition results in 5 frames after the consonant interval candidate are used, and the first
If the phoneme is recognized in the first place, it is given 2 points, and if it is recognized in the second place, it is given 1 point.The phoneme with the highest score is counted for each vowel in the five frames and is taken as the recognition result. (In Figure 2, /e/ has the highest score.) Using the method described above, the magnitude of the characteristic parameters low-frequency power and high-frequency power dip is obtained by obtaining samples from a large number of audio data in advance, and comparing them to each of the four phonemes. Create a discriminant diagram that discriminates into groups. An intermediate area is set at the boundary of each phoneme group in this discriminant diagram, and when the intermediate area is entered, consonant recognition is performed by matching with standard patterns of both adjacent phoneme groups. Further, a subsequent vowel is determined from the vowel recognition results for each frame, and a discriminant diagram for that vowel is applied.

Effects of the Invention As described above, the present invention applies the magnitudes of the low-frequency power dip and high-frequency power dip to a discriminant diagram, classifies them into four phoneme groups and an intermediate region, and matches them with the phoneme standard pattern to be recognized. This allows for highly accurate consonant recognition. Since this uses both low-frequency power and high-frequency power as parameters, it is effective for both voiced consonants that tend to have a power dip in the high-frequency power and voiceless consonants that tend to have a power dip in the reduced power. This is because it is working. By setting an intermediate region in the discriminant diagram and recognizing consonants classified in this region as both phoneme standard patterns of adjacent phoneme groups, it is possible to recognize consonants that are incorrectly classified near the boundaries of the discriminant diagram. It is now possible to improve the rate. Furthermore, since a discriminant diagram is created for each subsequent vowel, it is possible to create a discriminant diagram with higher accuracy. As described above, by using the method of the present invention, consonants can be recognized with high accuracy,
The effect is also great.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of a function that embodies a consonant recognition method in an embodiment of the present invention, FIG. 2 is an explanatory diagram of a power dip in this embodiment, and FIGS. 3 to 7 are in this embodiment The power dip distribution diagrams of voiceless plosives, voiced plosives, nasals, and voiceless fricatives in FIGS. 8 to 12 are /ra/ and /r +/ in this example. Figure 13 is a block diagram of a conventional word recognition system, and Figure 14 is a diagram showing the temporal change and rate of change in the power uttered in words /, u/, /, e/, /ro/, respectively. FIG. 2 is an explanatory diagram of a conventional consonant segmentation method. 10... Power dip detection section, 11...
...Power dip size extraction part, 12...
Vowel recognition unit for each frame, 13...Subsequent vowel recognition unit, 14...Discriminant map selection unit, 15...
・Subsequent vowel side discriminant diagram storage unit, 16...Major classification judgment unit, 17...Standard pattern selection unit, 18...
... Standard pattern storage section, 19 ... Consonant recognition section. Name of agent: Patent attorney Toshio Nakao and 1 other person No. 1
Figure 7 PL → Figure 8 Figure 91m Figure 10 Figure 110 Figure 12

Claims (4)

[Claims] (1) In a speech recognition method characterized by phoneme recognition, the low-frequency power and high-frequency power of the speech spectrum are obtained, the magnitude of the power dip caused by each temporal change is extracted, and these are analyzed for each subsequent vowel. Consonants are classified into several consonant groups and intermediate regions by applying a discriminant diagram created in advance, and the consonant groups are matched with the phoneme standard pattern of each consonant group, and the intermediate regions are matched with the adjacent phoneme pattern. A consonant recognition method characterized by performing matching with standard phoneme patterns for both consonant groups. (2) Consonant recognition according to claim 1, characterized in that the magnitude of the power dip is extracted as the magnitude of the temporal change speed when the power changes from the consonant to the subsequent vowel. Law. (3) The method for creating a discriminant diagram is to calculate the distribution from the size of the power dip that appears in each consonant group based on a large amount of data in advance, and express the discriminant results as a discriminant diagram for all the input data expected in advance. 2. The consonant recognition method according to claim 1, wherein the consonant recognition method is used as a consonant recognition method. (4) The consonant recognition method according to claim 1, wherein an intermediate region is set near the boundary of the discriminant diagram to reduce errors caused by the discriminant diagram.