JPH0114600B2 - - Google Patents

Description

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

Configuration of conventional examples and their problems The pattern matching method is often adopted as the operating principle of automatic speech recognition systems that have been previously researched or published. This method memorizes standard patterns for all types of words that need to be recognized in advance, and compares them with unknown input patterns to calculate the degree of matching (hereinafter referred to as similarity). The word is calculated and determined to be the same word as the standard pattern that yields the maximum match. In this pattern matching method, standard patterns must be prepared for all words to be recognized, so if the speaker changes, a new standard pattern must be input and stored. Therefore, in cases where there are hundreds of types of words to be recognized, such as the names of cities across Japan, it takes a huge amount of time and effort to pronounce and register all the types of words. It is expected that the memory capacity will also be enormous. Furthermore, there is a drawback that the time required for pattern matching between the input pattern and the standard pattern increases as the number of words increases.

On the other hand, input speech is divided into phoneme units and recognized as combinations of phonemes (hereinafter referred to as phoneme recognition).
The method of determining similarity with a word dictionary written in phoneme units has the advantage that the memory capacity required for the word dictionary is significantly reduced, the time required for pattern matching is shortened, and the contents of the dictionary can be easily changed. I have it. An example of this method is described in ``Word speech recognition system using the outline form of the speech spectrum and its dynamic characteristics'' by Miwa et al., Journal of the Acoustical Society of Japan 34 (1978).

A block diagram of a word recognition system using this method is shown in FIG. First, the voices of multiple speakers are analyzed in advance by the acoustic analysis unit 1 using a filter bank for each 10ms analysis interval, and the feature parameters are determined by the feature extraction unit 2 based on the obtained spectrum information. . From this feature parameter/
A standard pattern is created for each phoneme or phoneme group, typified by vowels such as a/, /o/, and consonants such as /m/, /b/, and is registered in the standard pattern registration section 5. Next, the input speech of an unspecified speaker is similarly analyzed by the acoustic analysis section 1 for each analysis section, and the feature extraction section 2 obtains feature parameters. Using these characteristic parameters and the standard pattern of the standard pattern registration section 5, the segmentation section 3 performs a separation operation between vowels and consonants (hereinafter referred to as segmentation). Based on this result, the phoneme discriminator 4 compares the phoneme with the standard pattern in the standard pattern registration unit 5 to determine the phoneme corresponding to the standard pattern with the highest degree of similarity in that section. Finally, the time series of phonemes created as a result (hereinafter referred to as phoneme series)
is sent to the word recognition unit 6, and the word corresponding to the item having the highest degree of similarity to the word dictionary 7 similarly expressed as a phoneme sequence is output as a recognition result.

As can be seen from the overall operation described above, if the segmentation unit 3 makes a mistake in segmentation, a phoneme that should be present may be overlooked (dropped phoneme), or a different phoneme may be inserted where there is actually no phoneme. This results in the incorporation of phonemes (addition of phonemes). When these errors occur, when a word is expressed as a phoneme sequence, the omission or addition of a phoneme causes it to resemble another completely unrelated word, increasing the risk of misrecognition. .

As described above, segmentation is the most important task in a method of word recognition based on phoneme recognition, and the performance of a word recognition system is greatly influenced by the accuracy of segmentation. Conventionally, as a parameter for performing segmentation, the temporal movement of power information in the spectrum of the entire audio signal band is used, and segmentation is performed based on the presence of power dips as shown in Figure 2. I was getting used to it. That is, by utilizing the fact that the power of the vowel part is greater than the power of the consonant part, a part where the dip size D is larger than the threshold value θ _D (D>θ _D ) is defined as a consonant section. This method had the following two problems.

(1) There are phonemes for which the existence of dips is not clear in the information of all bands, and the accuracy is not good. (especially/
r/, /η/, /h/, /m/, /n/, etc.) (2) The depth D is expressed as the difference between the power of the left and right vowels. Therefore, if the power movement in the vowel interval is not simple, it is difficult to directly determine the magnitude of the dip.

OBJECTS OF THE INVENTION The present invention solves these problems and aims to perform segmentation within words with high accuracy.

Structure of the Invention In Japanese, words and sentences are usually composed of vowels and consonants that are alternately combined, and consonants and other consonants, except for plosives, are not consecutive. Therefore, when recognizing Japanese speech, if vowels and consonants can be separated with high accuracy, it will greatly contribute to improving the recognition rate. The present invention uses both the low-frequency power and the high-frequency power of the speech spectrum as information used for segmentation, and uses the power dips caused by the temporal movement of each to accurately detect consonant intervals. The aim is to improve the accuracy of segmentation within words.

DESCRIPTION OF EMBODIMENTS FIG. 3 shows a typical spectrum pattern of a phoneme. a is a five-vowel sound, b is a nasal sound, the droning part of a voiced plosive, and c is a voiceless consonant.
As is clear from these figures, the power of a is relatively concentrated in the middle range, the power of b is concentrated in the low range, and the power of c is concentrated in the high range. In addition to these, Ryuon /r/
There are also phonemes whose spectrum is greatly influenced by the preceding and following phonemes, such as the nasal sound /η/. Taking these matters into consideration, the power in the high range is effective in distinguishing between vowel group a and voiced consonant group b, and the power in the low range is effective in distinguishing between vowel group a and voiceless consonant group c. It can be seen that the magnitude of power is effective.

Based on the above knowledge, in this example, the low frequency power obtained by smoothing the output of the 250Hz-600Hz bandpass filter is used as the segmentation parameter for the low frequency information.
For high frequency information, the high frequency power obtained by smoothing the output of a 1500Hz-4000Hz bandpass filter is used. By using both low-frequency power and high-frequency power as in this example, compared to the conventional example that uses only full-range power, the
m/, /n/, /η/, /r/, /h/, /z/
It was possible to obtain a large power dip compared to the current value, and the detection accuracy was improved.

However, in order to calculate the absolute value of the power dip, it is necessary to use a wide range of information before and after the dip, so in the conventional method, the procedure is complicated and there are many detection errors. In this embodiment, a simple and accurate dip detection method that takes into account the limitations of the vocalization mechanism is adopted.

Speech production is a combination of muscle movements, including the lungs and trachea, which control exhalation, the vocal cords, which produce voiced sounds, and the articulator tubes, which determine phoneme. Therefore, the movement of vocal power is constrained by the movement of the muscles of the vocal tube. For this reason, the temporal rate of change in voice power remains within a certain range, although there are fast-moving sounds such as plosives and slow-moving sounds such as semi-vowels. Therefore, there is no practical problem even if the magnitude of the ID step is replaced by the amount of change in power within a unit time. The dip detection method based on this idea will be specifically described below.

FIG. 4 explains the method. Power information is calculated for each frame (one frame is 10 msec) using logarithmically transformed power information. Let P(i) be the logarithmic power information in the i-th frame (i=1 to i _nax , where i _nax is the final frame of the voice section).
FIG. 4a shows an example of the temporal movement of the logarithmic power information P(i) in the series of vowels, consonants, and vowels. In this figure, in addition to large dips in the consonant interval, small dips due to small fluctuations in power are superimposed. As mentioned earlier, the fine dips are not caused by the muscle movements necessary for vocalization, so they are removed by smoothing. What has been removed is shown in Figure 4b. The power information (i) after smoothing is (i)={P(i-1)+2×P(i)+P(i+1)}/4. Next, the difference value P _D of the power information after smoothing is calculated by the following equation, and the temporal change of the power information is determined (FIG. 4c).

P _D (i)=(i+1)-(i-1) That is, P _D represents the temporal movement of the amount of change every 20 msec. P _D reaches its minimum value at the downward inflection point of the power dip, and reaches its maximum value at the rising inflection point. For the reason mentioned above, the size of the dip is the size P between the maximum and minimum values of P _D
Replace it with Also, the duration of the dip is P _D
Let L be the time from the minimum value to the maximum value.

Low frequency information ( _PL ) mentioned earlier as power information
By using both the low-frequency information (P _H ) and the high-frequency information (PH) and applying the method explained in FIG. 4 to each of them, it is possible to obtain the dip due to the low-frequency information and the dip due to the high-frequency information, respectively. Among these dips, only those that satisfy the condition L≦L _nax are selected as consonant candidates. Generally, consonant intervals are 150 msec (L _nax = 15) or less, excluding /s/ and pellicles, so this condition is included. /s/ and the cursive sound can be detected by other methods.

Among the voice segments found as consonant candidates, there are those found using only low-frequency information ( _PL ) and those found only using high-frequency information ( _PH ). Furthermore, these consonant candidate sections are of two types: real consonant sections and non-true consonant sections (addition of consonants).
Next, a method for separating consonant intervals and consonant additions from consonant candidate intervals will be described.

Let the magnitude of the change in dip obtained from the low frequency information _PL and the high frequency information _PH be p _l and _ph , respectively. Statistically, dips appear more prominently in true consonant intervals than in consonant additions, so both or one of p _l and p _h takes a large value. For example, the phoneme /b/ has a large value for both p _l and p _h , /h/ has a large value only for p _l , and /m/ has a large value for p _h
only becomes larger. On the other hand, for dips due to the addition of consonants, both p _l and p _h have relatively small values. Taking these characteristics into consideration, a discriminant diagram in the p _{l -ph} space is used to accurately and efficiently distinguish between consonants and additions.

FIG. 5 is an example of a discriminant diagram. In the figure, the area inside the shaded area is the addition area, and the area outside the area is the consonant area. however
p _l and p _h are converted into integers and normalized. The discriminant diagram uses a large amount of data that has been obtained by visual inspection of segmentation in advance, and takes into account both the probability of being correctly recognized as a consonant and the probability of addition, in order to optimize the results. It has been decided.

Next, a method of determining a consonant interval using a discriminant diagram will be explained using the example shown in FIG. Figure 6a
is the case when only the dip of p _l appears, and the size is p _l =10. When this is applied to the discriminant diagram in FIG. 5, (10, 0) is an additional area, so it is not a consonant interval. b is p _l =7, p _h =8,
Located in the consonant region. In this case, the logical sum of both the intervals p _l and p _h is taken as the consonant interval (depending on the phoneme, it may not be the logical sum). c is an example of an interval in which only _ph exists, and (0, 12) is located in the consonant region on the discriminant diagram. In this case, the p _h interval is directly used as the consonant interval. Although d has dips in both p _l and p _h , it is located in the addition region on the discriminant diagram, so it is treated as addition.

This example was evaluated using 212 words uttered by 10 men and 10 men. This word set is a set for evaluation in which consonant sections are labeled in advance by visual inspection. The results obtained when this example was applied were compared with the labels, and the percentage of correct segmentation was evaluated. The results (correct answer rate) are shown below.

/r/: 94.7%, /h/: 94.8%, /z/: 98.7
%, /b/: 99.5%, /d/: 99.7%, /η/: 91.3
%, /m/: 85.7%, /n/: 85.7% On the other hand, the probability (addition rate) of a consonant being mistakenly added to a vowel interval is 6.9%.

Comparing this result with the conventional method (method that uses full-band spectral power and detects dips using a threshold value), it is found that /r/, /h/, /η/ has a decrease of several percent, /
There is an improvement of about 1% in b/ and /d/. Also/
Although dips cannot be detected in m/ and /n/ with full band power, they can be detected in this embodiment. The addition rate is almost the same.

In this way, this embodiment can detect consonants in words (especially /r/, /η/, /
h/, etc.) can be performed with high precision.

Effects of the Invention As described above, according to the present invention, segmentation accuracy is improved by using both low-frequency power information and high-frequency power information as parameters.

Furthermore, by utilizing the temporal movement and duration of the power dip, the presence of the dip can be easily detected.

Furthermore, by using the magnitude of the movement of the power dips in both the low and high ranges and applying them to the discriminant diagram, the presence of consonants can be detected with high accuracy.

[Brief explanation of drawings]

Figure 1 is a block diagram of a conventional speech recognition system, Figure 2 is a diagram explaining a conventional method of detecting consonants using power information, and Figures 3a to 3c show examples of vowel and consonant spectra. Figures 4a to 4c are diagrams illustrating a method for detecting power dips according to the present invention, and Figure 5 shows a method for determining consonants and additions based on the respective sizes of low-frequency power dips and high-frequency power dips. FIG. 6 is a diagram showing an example of a method for determining consonant intervals.

Claims (1)

[Claims] 1. As information used for segmentation in speech recognition, the low-frequency power and high-frequency power of the voice spectrum are used together, and the power dip caused by the temporal movement of the respective powers is calculated. A speech segmentation method comprising: detecting a consonant candidate interval using the method of detecting a consonant candidate interval; and detecting a consonant interval from among the consonant candidate intervals. 2 Find the temporal change rate of each of the low-frequency power and the high-frequency power, detect consonant candidates based on the local maximum value, local minimum value, and time length of the temporal change rate, and calculate the local maximum value and local minimum value for the consonant candidate. A patent characterized in that a consonant interval is detected from a consonant candidate interval by regarding the value between the values as the magnitude of the power dip and applying the magnitude of the power dip of each of the low-frequency power and the high-frequency power to a two-dimensional discriminant diagram. A speech segmentation method according to claim 1.