JPH026078B2 - - Google Patents

Description

[Detailed description of the invention]

INDUSTRIAL APPLICATION FIELD The present invention relates to a method for segmenting word-initial consonants used in speech recognition devices and the like. Conventional configurations and their problems The pattern matching method has often been adopted as the operating principle of speech recognition systems that have been 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 provides the maximum similarity. In this pattern matching method, standard patterns must be prepared for all words to be recognized, so if the speaker changes, it is necessary to input and memorize a new standard pattern. Therefore, in cases where hundreds of types of words are 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 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. have. 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 section 1 using a filter for each analysis interval of 10 ms, and the feature parameters are obtained by the feature extraction section 2 based on the obtained spectrum information. From this feature parameter /a/, /
A standard pattern is created for each phoneme or phoneme group, typified by vowels such as o/ and consonants such as /n/ and /b/, and is registered in the standard pattern registration section 3. 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. Segmentation is performed in the segmentation unit 4 using these feature parameters and the standard pattern stored in the standard pattern registration unit 3. Based on this result, the phoneme discriminating unit 5 compares the phoneme with the standard pattern in the standard pattern registration unit 3 to determine the phoneme corresponding to the standard pattern with the highest degree of similarity as the phoneme in that section. Finally, the time series of phonemes created as a result (hereinafter referred to as the phoneme series) is sent to the word recognition unit 6, and the word corresponding to the item with the highest similarity to the word dictionary 7 similarly expressed in the phoneme series is recognized. Output as result. Next, a method of segmenting initial consonants in segmentation 4 will be described. Conventionally, the segmentation method for word-initial consonants takes advantage of the fact that the spectrum of voiced consonants at the beginning of words is similar to the spectrum of nasal sounds, and performs phoneme recognition on five vowels and a nasal sound in each frame. Judgment was based on the presence or absence of phoneme recognition results. For example, /ma/ at the beginning of a word often appears as the phoneme sequence /NNNNAAAA/ when phoneme recognition is performed for each frame, and among these /
A method was used to distinguish between /m/ and /i/ by making the NNNN/ part a voiced consonant. In addition, the presence of voiceless consonants with short speech sounds was distinguished by finding a phenomenon in which the slope of the spectrum changes markedly. For example, /pa/ at the beginning of a word.
Because the slope of the spectrum often changes rapidly at the transition from /p/ to /a/,
He used this to distinguish between /p/ and /a/. However, the above method cannot always detect the presence of consonants, and often misses the initial consonant (hereinafter referred to as consonant omission), or makes the mistake of treating the beginning of a word as a consonant interval even though it begins with a vowel (hereinafter referred to as consonant addition). There are many. It is in voiced consonants/
This is because consonants such as r/, /b/, /d/, etc. do not necessarily exhibit nasality. Furthermore, voiceless consonants with short durations such as /p/ and /t/ do not necessarily exhibit a significant temporal change in the spectral slope. OBJECTS OF THE INVENTION The present invention solves the above-mentioned drawbacks, and it is an object of the present invention to provide a highly accurate segmentation method for word-initial consonants with fewer omissions and additions of word-initial consonants. Structure of the Invention In order to achieve this object, the present invention has the following four features.
Two methods ●The first method of detecting word-initial consonants by capturing temporal fluctuations in the low-frequency power and high-frequency power of the speech spectrum at the beginning of words.●The second method of detecting word-initial consonants using nasality at the beginning of words. ●Third method of detecting word-initial consonants using asexual consonance at the beginning of words ●The frame in which the vowel spectrum appears temporally stable for the first time from the beginning of the word is used as the reference frame, and the spectral pattern of this reference frame and the word-initial A fourth method of detecting a word-initial consonant by comparing the spectral patterns of each frame from to a reference frame is applied in an arbitrary order, and when a consonant is detected, no subsequent methods are applied; This invention provides a method for segmenting word-initial consonants, which is characterized by performing segmentation of word-initial consonants based on detection results. DESCRIPTION OF EMBODIMENTS An embodiment of the initial consonant segmentation method of the present invention will be described below. This example uses a first word-initial consonant detection method that captures temporal fluctuations in the low-frequency power and high-frequency power of the speech spectrum at the beginning of words, and performs voiced/unvoiced judgment for each frame of the speech interval, and uses the results. The second word-initial consonant detection method used, the third word-initial consonant detection method that performs phoneme recognition on five vowels and nasals in each frame and uses the results, and the spectral pattern of the reference frame and the word-initial detection method. The fourth word-initial consonant detection method, which detects by comparing the spectral patterns of each frame up to the reference frame, is applied in this order, and when a consonant is detected, the subsequent methods are not applied, and the detection result is This method performs segmentation of initial consonants based on .
Each word-initial consonant detection method will be explained in detail below. First, a first word-initial consonant detection method that uses low-frequency and high-frequency power information will be described. The reason why low-frequency power and high-frequency power are used together in this embodiment is that characteristics of voiced consonants tend to appear in high-frequency power, and characteristics of voiceless consonants tend to appear in low-frequency power. The low-frequency power is obtained by passing the audio signal through a low-frequency bandpass filter, obtaining a power value for each frame, and smoothing it. Also, the high frequency power can be obtained in the same manner using a high frequency band filter. FIG. 2 shows an example of a temporal change in low frequency or high frequency power at the beginning of a word. When the beginning of a word is mainly a plosive consonant, the temporal change in power value is plotted as shown in Figure 2a. This is because the power rises rapidly due to its plosive nature, and the part where it crosses with the following vowel becomes concave like a. b is the value obtained by differentiating the value of the temporal change in power of a. _P1 to _P3 indicate frame numbers of inflection points of a. Here, the frame number at which the voice section begins is set to 1. Here, like a and b
The differential values of P ₁ and P ₃ are positive, the differential value of P ₂ is negative, and P ₃
<m (m is a threshold value indicating a frame number), the period from the beginning of the word to _P3 is determined as the initial consonant section. Applying the above method to low frequency power and high frequency power, if the tendency of a appears in either one,
The section is determined to be a consonant. This method is effective for all plosives, since characteristics tend to appear in the low-frequency power of voiceless plosives and in the high-frequency power of voiced plosives. Next, a second word-initial consonant detection method using nasality at the beginning of a word will be described. The phoneme recognition of this embodiment is performed in each frame (for example, one frame is 10 msec).
). In this embodiment, phoneme recognition for each frame is performed using LPC cepstral coefficients by comparison with a standard pattern of each phoneme prepared in advance. The standard pattern is 5 vowels (/a/, /i/, /u/, /e/, /o/),
Nasals (represented by /N/) and voiceless consonants (represented by /s/) were used. In this way, for each frame, the phoneme with the highest degree of similarity (first candidate phoneme) and the two
The phoneme with the highest degree of similarity (second candidate phoneme) is found. The sequences in which the first candidate phonemes and second candidate phonemes for each frame are arranged in numerical order are defined as a first candidate phoneme time series and a second candidate phoneme time series. When the above phoneme sequence is viewed in order from the beginning of the word, if /N/ continues for a certain number of frames or more (for example, 4 frames or more) including the first candidate or second candidate phoneme sequence, this section is determined to be a consonant section. For example, when /ma/ is phoneme recognized frame by frame, if the phoneme recognition result is as shown in Figure 3, there are 5 consecutive frames of /N/ including the first and second candidates, so 1 to 5
The period up to the frame is taken as the word-initial consonant section. This method is particularly suitable for /m/, /n/, /b/, /d/, /
It is effective for g/etc. Next, a third word-initial consonant detection method that utilizes voiceless consonance will be described. Segmentation of voiceless consonants at the beginning of words can be performed accurately by using the voiced/unvoiced determination results performed for each frame. There are methods for determining the presence/absence of voice, using zero-crossing waves, the slope of the spectrum, the value of the first-order autocorrelation coefficient, etc., but any method may be used. In this embodiment, the determination is made by comparing with a standard pattern for voiced/unvoiced determination. Here, when the unvoiced state continues for a certain number of frames or more from the beginning of the word (for example, 4 frames or more), this section is determined to be a consonant section. This method is valid for all voiceless consonants. Next, there are phonemes such as /z/ where the first half of the phoneme is voiced and the second half is unvoiced, so after a voiced frame for a certain length of time from the beginning of the word,
When unvoiced frames continue, this interval is considered a consonant. For example, /zu/ is judged frame by frame.
When VVVVUUUVVVV/ (where /V/ represents voice and /U/ represents unvoiced), the initial consonant is defined as the end of the unvoiced frame. Next, a fourth word-initial consonant detection method will be described in which the spectral pattern of the reference frame is compared with the spectral pattern of each frame from the beginning of the word to the reference frame. In this embodiment, LPC cepstral coefficients C ₁ to _Co (where n is a positive integer) are used as parameters representing the characteristics of the spectral pattern.
As a method of selecting a reference frame in which the speckle appears stably, the m-th frame from the beginning of the word (m=7 in this embodiment) is fixed. This is because consonants of relatively short duration are mainly detected. Equation (1) is used as a method for comparing spectrum patterns between two frames. f(i, j)= _o 〓 ^l=1 (C _l (i)−C _l (j)) ² ………(1) In equation (1), C _l (i) is l at the i-th frame from the beginning of the word. It represents the th LPC cepstral coefficient. Similarly, C _l (j) represents the l-th LPC cepstrum coefficient in the j-th frame. f
The larger the value of (i, j), the more different the spectral patterns of the two frames are. Using this equation (1), calculate f(i, m) (where 1≦i≦m−1) between the reference frame and each frame from the beginning of the word to the reference frame, and set the maximum value to f _nax . The presence or absence of a word-initial consonant is determined depending on whether the value of f _nax is larger or smaller than a certain threshold value. When detected using this method, the word-initial consonant interval is f
The period up to the frame in which the change in the value of (i, m) is the largest is defined as a consonant section. An example is shown in FIG. The horizontal axis is the time axis when the frame number at the beginning of the word is 1, and the vertical axis is f(i, m) when the frame number m is the reference (1≦i≦m−
1). In the figure f(i, m)
Since the maximum value f _nax = f (1, m) is larger than the threshold θ, it means that the initial consonant has been detected,
The initial consonant section is set up to frame number 3 where the change in f(i, m) is the largest, and segmentation is performed using frames 1 to 3 as the initial consonant section l. The present example and the conventional method were compared using data (approximately 2100 words) in which 10 men each uttered 212 words. The table shows the detection rate of word-initial consonants using the conventional method and the detection rate of word-initial consonants using the method of this embodiment. The method of this embodiment uses the first and fourth initial consonant detection methods to detect plosives (/p/, /t/, /
k/, /b/, /d/) is improved, and the second
Voiceless consonants (/h/, /
s/, /c/), and the third word-initial consonant detection method improves the detection rate for all voiced consonants. On average, the initial consonant detection rate has improved from 85% to 96% compared to the conventional method. In addition, although the beginning of a word starts with a vowel, the rate of mistakenly determining that it is a consonant (consonant addition) is 24 in the conventional example.
%, but with the method of this example, it has decreased to about 20%. Note that the order in which the first to fourth word-initial consonant detection methods are applied can be selected arbitrarily; in any case, when a consonant is detected, the following methods are not applied, and the initial consonant is detected based on the detection results. A similar effect can be obtained by performing segmentation.

【table】

[Table] Effects of the Invention As described above, the present invention provides a first method that captures temporal fluctuations in the low and high frequency power of the speech spectrum at the beginning of a word, a second method that uses nasality at the beginning of a word, Applying a third method of using asexual consonance at the beginning of a word and a fourth method of comparing the spectral pattern of the reference frame with the spectral pattern of each frame from the beginning of the word to the reference frame in any order, This method provides a segmentation method for word-initial consonants, which is characterized in that when a consonant is detected, the subsequent methods are not applied, and the word-initial consonant is segmented based on the detection results. This method has the advantage of improving the addition ratio of consonants to vowels and allowing highly accurate segmentation of initial consonants.

[Brief explanation of drawings]

FIG. 1 is a block diagram of a conventional speech recognition system, FIG. 2 is a diagram showing a reference frame detection method based on power information of the initial consonant segmentation method according to an embodiment of the present invention, and FIG. FIG. 4 is a diagram showing a reference frame detection method based on the phoneme recognition results of the same method, and FIG. be. 1... Acoustic analysis section, 2... Feature extraction section, 3...
Standard pattern registration unit, 4... Segmentation, 5... Phoneme discrimination unit, 6... Word recognition unit, 7...
...word dictionary.

Claims (1)

[Claims] 1 The first method detects word-initial consonants by capturing temporal fluctuations in the low-frequency power and high-frequency power of the speech spectrum at the beginning of words, and the second method detects word-initial consonants using nasality at the beginning of words. , the third method detects the initial consonant using the voiceless consonance at the beginning of the word, and the frame in which the vowel spectrum first appears temporally stable from the beginning of the word is taken as a reference frame, and the spectral pattern of this reference frame and the beginning of the word are A fourth method of detecting a word-initial consonant by comparing the spectral patterns of each frame from to a reference frame is applied in an arbitrary order, and when a consonant is detected, the subsequent methods are not applied, A method for segmenting word-initial consonants, which is characterized by performing segmentation of word-initial consonants based on detection results.