JPH0451036B2 - - Google Patents

Description

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

Configuration of conventional examples and their problems In recent years, there has been active development of speech recognition methods that use phonemes or syllables as basic units. In this method, segmentation of speech into units of phonemes or syllables is an important technical element for improving the speech recognition rate.

Conventionally, methods are known for segmenting phonemes or syllables that utilize the entire spectrum or band power. Here, as an example of a conventional method, a consonant segmentation method using the temporal movement of the spectral band power and the temporal concavity (power dip) of the power value will be described.

The conventional method will be described below with reference to the drawings. FIG. 1 is a functional block diagram of a conventional segmentation method. 1 is an AD converter that samples the input audio at 12KHz, and a band power calculation unit 2 uses a band filter to calculate high-frequency power and low-frequency power for each frame (10 msec). 3 is a power value buffer section which accumulates high frequency power and low frequency power to obtain time series information of power values. Then, the power dip extracting section 4 extracts the power dip from the time series information of the power values, and the phoneme section determining section 5 performs segmentation using the power dip section as a consonant section.

The conventional method takes advantage of the fact that since consonants have lower power than vowels, consonants tend to have power depressions. That is, in FIG. 2, when the time-series information of the power value indicated by a takes a smaller value than the surroundings, segmentation is performed as shown by b as a consonant from the fall of the power value to the vicinity of the rise. High range (1500~
4000Hz) power makes it easy to capture the dips of voiced consonants, and low-frequency (250-600Hz) power makes it easy to capture the dips of unvoiced consonants, so if you use both together, you can segment a wide range of consonants.

However, the drawback of the conventional example is that the detection rate is low for phonemes whose spectra are similar to vowels and have a small power difference from vowels, especially nasal sounds (/m/, /n/, /〓/, hatsu). It is. There is a method to compensate for this using nasality information (Hoshimi, Niyada: Segmentation method for word-initial consonants, Lectures on Phonetics, 1982)
(March 2013), nasal information is susceptible to noise and articulatory coupling, making stable segmentation impossible.

OBJECTS OF THE INVENTION The present invention eliminates the above-mentioned drawbacks of the prior art and provides a speech segmentation method that accurately performs segmentation of all kinds of phonemes, including nasal sounds.

Structure of the Invention In order to achieve the above object, the present invention calculates the degree of similarity between the feature parameters and the stationarity pattern for each frame, and calculates the degree of similarity between the feature parameters and the stationarity pattern for each frame. The present invention provides a method for carrying out this process.

Description of examples An embodiment of the present invention will be described below.

The present invention is based on the principle of capturing temporal changes in input parameters by comparing input parameters with stationarity patterns. First, a method for creating a temporal stationarity standard pattern will be explained. The stationarity pattern is determined by using many samples using multiple frames (m frames, m = 3 in this example) of temporally stationary parts of the audio signal, such as the center of vowels and vowels. create. Let n be the number of feature parameters per frame.
In this embodiment, low-order parameters (C ₀ to _{C 4} ) of LPC cepstrum coefficients are used as feature parameters. Therefore, the number of feature parameters n=
It is 5.

A feature parameter vector C is created by arranging m×n (15) parameters as follows.

C=(C ₀ ¹ , C ₁ ¹ ,...C ₄ ¹ ,C ₀ ² ,C ₁ ² ,...C ₄
² , C ₀ ³ , C ₁ ³ ... C ₄ ³ ) (Equation 1) However, in C ^j/i , i is the order number and j is the frame number. For convenience, C is written as follows.

C=(C ¹ , C ² , C ³ . . . C ¹⁵ ) (Formula 2) The average value μ and the variance-covariance matrix W of C are calculated using many samples. Let the elements of μ be μ _i and the elements of W be W _i,j . When the number of samples is n, μ _i =1/N _N 〓 ^K=1 C ^i/k ...(Formula 3) W _i,j =1/N-1 _N 〓 ^K=1 (C ^i/k −μ _i ) (C ^i/k −μ _j ) ...(Equation 4) A stationarity pattern (standard pattern) can be created.

Next, a method of calculating the similarity between the input feature parameters and the stationarity pattern will be explained.

Arrange the feature parameters (LPC cepstral coefficients) of the input speech in time series as in (Equation 1),
Let this be X.

X=(X ₁ , X ₂ , X ₃ ...X ₁₅ ) (Equation 5) The probability density P for the flatness pattern of X is expressed by the following equation.

P=(2π) ^-15/2 |W| ^-1/2 exp{-1/2(X- μ)'W ^-1・(X-μ)} ...(Formula 6) However, '' indicates transposition. represent

If we take the logarithm of (Equation 6) and double it to make it L, then L=-(X-μ)'・W ^-1・(X-μ)+A
(Formula 7) A is a constant A=2・log {(2π) ^-15/2・|W| ^-1/2 }
...(Formula 8).

For the voice section, calculate X while shifting one frame at a time, and then calculate the similarity using (Equation 7). In the stationary part, the value (similarity) of (Equation 7) becomes large, and the spectrum If there is a change or a change in power, the value of (Equation 7) will be small. Portions with low similarity correspond to phoneme boundaries or word boundaries, so segmentation can be performed by capturing these.

FIG. 3 shows, as an example, the change b in the degree of similarity when uttering ``Osama'' (/oosama/).
For reference, the figure shows the power change a according to the conventional example.
The phoneme label c, which was added by visual inspection, is also attached. According to FIG. 3, the change b in the degree of similarity forms minimum values at word boundaries and phoneme boundaries, thereby making it possible to easily segment phonemes. Comparing with the visual label c, it can be seen that the section can be detected successfully. On the other hand, in the power change a according to the conventional example, /s/ can be detected, but the nasal sound /m/ cannot be detected.

Figure 4 shows another example.
This is the case when you say inaho/). In this case as well, minimum values appear at word boundaries and phoneme boundaries in the similarity change b, and segmentation including nasal sounds is performed accurately. In the conventional case of power change a, segmentation is impossible.

Next, FIG. 5 shows a functional block diagram for realizing the method described above.

In FIG. 5, the AD converter 1 has the same function as the conventional example, so the explanation will be omitted. Reference numeral 10 denotes an acoustic analysis section that analyzes audio signals, and in this embodiment,
LPC analysis method is used. The analysis window is a Hamming window, the frame period is 10 msec, and the analysis order is
It is 15. 11 is a feature parameter extraction unit;
Power term C ₀ and four lower-order parameters (C ₁ to _{C 4} )
Calculate. Reference numeral 12 denotes a similarity calculation unit, which calculates the similarity between the input feature parameter and the stationarity pattern using (Equation 7).

13 is a stationarity pattern storage unit, (Equation 3),
It contains the values of (Formula 4) and (Formula 8). The time series buffer 14 stores similarity information as a time series. The phoneme interval determination unit 15 detects a portion where the degree of similarity is small from the time information of the degree of similarity, and determines a phoneme interval as illustrated in FIGS. 3 and 4.

In this way, the segmentation method of this embodiment captures temporal changes in spectra at phoneme boundaries as similarity information, so it can accurately segment phonemes such as nasals whose power values are not much different from vowels. be able to. Furthermore, since the temporal change in similarity is used as a relative value (that is, by detecting the minimum value), it is less susceptible to the effects of noise and articulatory combination.

In addition, in the above example, the feature parameter is
Although we used LPC cepstral coefficients, it is also possible to use other characteristic parameters such as band spectral power, PARCOR coefficient, autocorrelation function, autocorrelation coefficient, etc. Further, the order of the LPC cepstrum coefficients does not need to be limited to C ₀ to C ₄ .
Further, although three frames are used in the above example, the number of frames is not limited to three frames if it is a plurality of frames.

Regarding distance measures, other statistical distance measures,
For example, Mahalanobis distance may be used. In this case, the phoneme interval determination unit 15 in FIG. 5 detects the maximum value and performs segmentation.

In addition, when creating a temporal stationarity pattern, it is said that it is created using samples from the center of vowels and syllables, but in reality, it is created by shifting one frame at a time in all voiced sound intervals or all speech intervals. (In general speech, there are more stationary parts than parts where the spectrum changes.) Effects of the Invention In summary, the present invention calculates the similarity between feature parameters and stationarity patterns for each frame. This method provides a method for segmenting phoneme intervals by calculating and capturing temporal information changes in similarity, and is useful for phonemes that cannot be accurately segmented using conventional methods, such as nasal sounds. This method can perform segmentation with high accuracy, and since changes in the spectrum are detected based on the relative values of similarity information, it has the advantage of being less susceptible to fluctuation factors such as noise and articulatory combination. Furthermore, since all similarity calculations are sum-of-product calculations, it has the advantage of being easy to implement in hardware.

[Brief explanation of the drawing]

Figure 1 is a functional block diagram for explaining the conventional voice segmentation method, Figure 2 is a diagram for explaining the conventional method for segmentation using power changes, and Figures 3 and 4. Figure 5 shows a specific example to demonstrate the effectiveness of the voice segmentation method in one embodiment of the present invention.
The figure is a functional block diagram for realizing this embodiment. 1... AD conversion section, 10... Acoustic analysis section, 11
...Feature parameter extraction unit, 12...Similarity calculation unit, 13...Stationality pattern storage unit, 14...Time series buffer, 15...Phoneme interval determination unit.

Claims (1)

[Claims] 1. Analyze input audio for each analysis section (frame) to obtain feature parameters, and statistically evaluate the similarity between the temporal pattern of the feature parameters and a standard pattern expressing temporal stationarity. Speech segmentation is performed by calculating the similarity using a distance measure, creating time-series information of the similarity for the speech interval, and detecting the boundaries of the speech using the temporal movement of the time-series information. A voice segmentation method characterized by performing the following. 2. The standard pattern expressing temporal constancy is
2. The speech segmentation method according to claim 1, wherein the speech segmentation method is comprised of an average value and a variance-covariance matrix using characteristic parameters of a plurality of frames of a large number of samples. 3 The feature parameters are LPC cepstral coefficients,
The speech segmentation method according to claim 1, wherein the segmentation method is one selected from band spectral power, PARCOR coefficient, and autocorrelation function. 4. The speech segmentation method according to claim 1, wherein the statistical distance measure is any one of probability density, log likelihood, or Mahalanobis distance. 5. The method according to claim 1 or 2, characterized in that the standard pattern of temporal stationarity is created using any one of a stationary part of speech, a voiced sound interval, and a whole speech interval. Audio segmentation methods.