JPH04122999A - Phoneme discriminating method - Google Patents

Abstract

PURPOSE:To obtain high efficient performance by obtaining static information and power variation of a voice as power vectors and variation in spectrum as time-series feature vectors, and further performing hierarchic recognition. CONSTITUTION:The power vectors and time-series feature vectors are calculated and stored as power vector data and time-series feature data, and a power vector code book 105 is generated with the power vector data. Then vector quantization is carried out by using the code book 105 to generate power code data. Time-series feature vector data corresponding to frames given the same power code number in the power code data are clustered to generate a time- series feature code book 108, and this operation is repeated as many times as the number of power codes. For the time-series feature vector data, the code book 108 determined by the corresponding power code is used for vector quantization, thereby finding the time-series feature code data. The code data and time-series data are used to generate a phoneme table 110 showing the corresponding relation between both the codes, thereby obtaining the high efficient performance.

Description

[Detailed Description of the Invention] [Prior Art] Conventionally, this type of technology is described in ``Evaluation of interspeech information on a frame-by-frame basis based on mutual information'', Institute of Electronics, Information and Communication Engineers, Speech Research Group, 5P-103 (1988). There is something I can disclose.

One of the methods currently being most actively researched in the field of speech recognition is a recognition method using phoneme identification processing. The phoneme identification process converts input speech into a series of phonemes (equivalent to the pronunciation symbols). After the speech is converted into a phoneme sequence in this way, it is converted into the most appropriate character string (sentence) using, for example, a word dictionary or grammar rules.

The advantage of performing phoneme identification is that by separating acoustic level processing and character string processing, it is possible to freely expand the range of words and sentence types to be recognized.

FIG. 8 is a block diagram showing the configuration of a device implementing a conventional voice identification method. In the figure, 201 is an audio input terminal, 202 is an analysis section, 203 is a power vector calculation section,
204 is a power vector 70 part, 205 is a power vector codebook, 206 is a feature vector VQ part, 207 is a feature vector codebook, 209 is a phoneme conversion part, and 210 is a phoneme table.

The voice input from the voice input terminal 201 is sent to the analysis unit 202.
can be converted into a feature vector representing the feature. As the feature vector i, in-band frequency components extracted by one group of band-pass filters having different center frequencies are used for each small time period called a frame.

5 i = (Sil Si2..., 5ij-1,
5ij) Here, i is the frame number and j is the bandpass filter number.

Furthermore, the analysis unit 202 simultaneously calculates the power Pi for each frame. To simplify the explanation, the frame number at the beginning of the audio is assumed to be O1, and the frame number at the end of the audio is assumed to be 1 (i-...-2, -1, 0, 1, 2,..., I, I
+1. I+2.・ ).

The power vector calculation unit 203 calculates a power vector Fi obtained by combining the powers Pi of adjacent frames.

p i = (Pi-n, Pi-n+1. =・, Pi
-1, Pi, -, Finn-1, Finn) where n is the width of the adjacent frame.

The power vector calculation unit 204 refers to the power vector codebook 205, performs vector quantization (VQ), and obtains a power code Ci.

Ci = argmin d (pi, y'') where d (pi, y'') represents the distance between vector i and vector y'''.

The power vector quantization process roughly classifies the input audio according to the power shape, and the result of the major classification is the power code Ci.
becomes. ” is the power code hector, and m is the power food number.

1"-(y"-n, y"-n+1..., y'-1,
y'O, y"L...y"'n-1. y”n)
(m=1.2. . . . , M) Here, M is the power vector codebook size.

The feature vector VQ unit 206 uses the feature vector codebook 20
In addition, the feature vector codebook 207 stores codebooks for each power food. In other words, there are codebooks corresponding to M power vector codebook sizes.

The feature vector quantization process performs detailed classification based on the detailed features of speech, contrary to the general classification of the power vector quantization process.

First, a feature vector codebook for power code Ci is selected from feature vector codebook 207. In other words, this corresponds to switching the detailed identification dictionary to an appropriate one in consideration of the major classification results. Quantization (VQ) is performed using this selected codebook. Now, if c=ci, the feature code Zi can be expressed by the following equation.

Z i = armin d (S i , ×(c)')
(2) Here, X(c)' is a 6-feature code vector corresponding to power code C, and r is a feature vector code.

x(c)'=(X(c)'+,X(c)',,-,
X(c)', -1°X(C)'J) <r=
1.2. -, R(c)) R(c) is the feature vector code size corresponding to the power code C.

The phoneme conversion unit 209 converts the phoneme symbol L from these Ci and Zi.
Convert to i. Various methods can be considered for this processing, but the simplest table lookup method will be explained here. First, an example of the configuration of the phoneme table 210 is shown in FIG. 10, where power code C1=1, feature code Z
According to this table, when i=1, the phonetic symbol Li is a. Also, Cf=2. In Zishi 3, it can be converted to Li-e.

In this way, the input speech can be converted into a phoneme symbol string.

[Problem to be solved by the invention]

In phoneme discrimination processing, high discrimination ability cannot be obtained unless the various keys of speech are effectively captured and discriminated. When humans hear voices, they not only use the static information of the voice, that is, the intensity of the sound at a certain moment, the difference in timbre (spectrum), but also
Various experiments have proven that the dynamic information of the voice, the intensity of the sound, and temporal changes in the spectrum are important keys.

In the conventional phoneme identification method described above, a feature amount called a power vector F is used to capture changes in power, which is one of the keys for phoneme identification. In addition, the static information of the spectrum (feature vector codebook spectrum) is also taken into consideration.

However, the conventional phoneme identification method described above does not take into account changes in the spectrum of speech, which is the most important key to identifying similar phonemes, and as a result, there is a problem in that sufficient phoneme identification ability cannot be obtained. .

The present invention has been developed in view of the above-mentioned points, and an object of the present invention is to provide a phoneme identification method that can obtain high identification performance by taking into account changes in the spectrum of speech, which is the most important key to identifying phonemes by similarity. With the goal.

[Means to solve the problem]

In order to solve the above problems, the phoneme identification method of the present invention employs the following means (a) to (g).

(a) Frequency analysis of the input voice and calculation of the feature vector, which is a vector of frequency components of the input voice, and the power representing the strength of the input voice at fixed time intervals called frames; (b) Adjacent A process of calculating a power vector obtained by combining the audio powers of frames; (c) a process of calculating a time-series feature vector obtained by combining the feature vectors of adjacent frames; (d) Let's identify (e) A process of creating an input power/temporal characteristic pattern and an input time-series feature pattern by performing the processes (a) to (c) for the input audio, and (e) creating a large number of input power patterns in advance for the input power pattern. A process of vector quantizing the audio data using a previously created power vector codebook to obtain an input power code; a process of vector quantizing using a vector codebook of time-series features to obtain an input time-series feature code; and <g) a process of finding the likelihood of a phoneme for each frame from the input power code and the input time-series feature code.

[Effect]

According to the present invention, the static information of the voice, that is, the power of the sound at a certain moment, the difference in the spectrum, as well as the static information of the voice and the power change, are expressed as a power vector, and the change in the spectrum is expressed as a time-series feature vector. Capture. Furthermore, by performing hierarchical discrimination using both of these features, using the power vector in the first stage and the time series feature vector in the second stage, it is possible to achieve more efficient and higher phonological discrimination than using each feature separately. performance can be obtained.

〔Example〕

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

FIG. 1 is a block diagram showing the configuration of an apparatus for implementing the phoneme identification method of the present invention. In the same section, 105 is a power A4 codebook, 106 is a time-series feature vector generation unit, 107 is a time-series feature vector codebook, 108 is a time-series feature vector codebook, 109 is a phoneme conversion unit, 110 is a phonological table.

Fig. 2 is a diagram showing an example of input audio signal power, Fig. 3 is a diagram showing an example of input audio signal bang, Fig. 4 is a diagram showing an example of temporal change in frequency spectrum, and Fig. 5 is a time series pattern of spectrum. FIG. 6 is a diagram showing an example of a power vector codebook, and FIG. 7 is a diagram showing an example of a time-series feature vector codebook.

An audio signal that can be input from the audio input terminal 101 in FIG. 1 is converted into a time series of feature vectors representing features in the analysis section 102. Possible methods for deriving feature vectors include using multiple bandpass groups with slightly different center frequencies, or using spectral analysis using FFT (fast Fourier transform), but here we will use a group of bandpass filters. Here is an example of a method. As the feature vector Si, the in-band frequency components extracted by a group of pass filters having different center frequencies are logarithmically transformed and extracted for each small time period called a frame.

Si = (Si1.Si2..., Si j-1,
5ij) Here, i is the frame number and J is the bandpass filter number. Furthermore, the analysis unit 102 simultaneously calculates the power Pi for each frame. The power Pi can be calculated using the following equation.

P i =3−'Σ5ij
(3) For simplicity, let us assume that the frame number at the beginning of the audio is 0 and the frame number at the end of the audio is 1. (Combine the powers Pi to calculate the power vector Pi.

pi = (Pi-n, Pi-n+1.-, Pi-1,
Pi, Pi+1. -, Pi+n-1, pi+n) quantization (VQ) I, , determine the power code Ci.

Ci = argim d (p i , y”)
(4) Here, d([pi, y“) is the vector i
represents the distance between and Hectorno°. In the power vector quantization process, input speech is roughly classified into power shapes as shown in FIG. 2, and the results of the rough classification become power codes Ci as shown in FIG. 3. yl is a power code vector, and m is a power code number.

y"= (y--n, y'-n+L +..., y"-1
,y"0.y"1.・-, y”n-1, ymn)
(m=1.2...,M) Here, M is the power vector codebook size.

The time-series feature vector generation unit 106 generates a time-series feature vector obtained by combining feature vectors Si of adjacent frames.
Generate i.

Ti = (+1 to Si-, +1 to Si-...,
5i-1, 5i, Si+1, Si+k) where k is the adjacent frame width.

In time-series feature vector quantization 107, vector quantization (VQ) of time-series feature vectors Ti is performed based on a time-series feature vector codebook 108 as shown in FIG. Further, the time-series feature vector codebook 108 has all the codebooks stored for each power code. In other words, there are codebooks corresponding to M power vector codebook sizes. Contrary to the power vector quantization process described above, the time-series feature vector quantization process performs detailed classification based on the detailed features of the voice and, in particular, changes thereof.

A time-series feature vector codebook corresponding to the previous power code Ci is selected from the time-series feature vector codebook 108.

In other words, this corresponds to switching the detailed identification dictionary to an appropriate one in consideration of the major classification results. Quantization (VQ) is performed using this selected codebook.

Now, assuming c-Ci, the feature code Zi is the following formula % Formula % Here, (c)' is the time-series feature code number for the power code C.

u(c)'-(x(c)'+, x(c)'x,
+..., x (e)'+th++)1-1+ X (
C)'(t*+1)+ )(r=1.2..., R(c
)) R(c) is the time-series feature vector codebook size corresponding to the power code C.

The phoneme conversion unit 109 converts this power code C11 characteristic code Zi into a phoneme symbol Li. Various methods can be considered for this process, but the simplest table lookup method will be explained here. In the phoneme table 110 shown in FIG. 9, power chord C4-1
, when the feature code Zi=1, the phonetic symbol Li is a. Further, when the power code C3=2 and the feature code Zi=2, the phonetic symbol Li becomes e.

In this way, the input speech is converted into a phoneme symbol string.

Note that there are various possible methods for creating the phoneme table 110, and one method will be briefly explained in bullet points.

(a) Calculate a power vector and a time-series feature vector for a large number of audio data in advance, and store them as power vector data and time-series feature data, respectively.

(b) Clustering the power vector data and creating a power vector codebook.

(C) Vector quantize the power vector data using a power vector codebook to create power code data.

(d) The process of clustering time-series feature vector data corresponding to frames assigned the same power code number in the power code data and creating a time-series feature codebook is repeated for the number of power codes.

(e) Vector quantize the time-series feature vector data using a time-series feature vector codebook determined from the corresponding power code to obtain time-series feature food data.

(f> A phoneme table 110 is created that shows the correspondence between the two foods from the phoneme food data given to the audio data in advance through inspection, etc., and time-series feature data.

〔Effect of the invention〕

As explained above, according to the present invention, static information of audio,
In other words, not only the power of the sound at a certain moment and the difference in the spectrum, but also the static information and power changes of the voice are captured by the power vector, and the IT of the spectrum is captured by the time series feature vector.
Furthermore, by performing hierarchical discrimination using both of these features, using the power vector in the first stage and the time series feature vector in the second stage, it is possible to achieve more efficient and higher phonological discrimination than using each feature separately. An excellent effect can be obtained in that the method can be obtained.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing the configuration of an apparatus implementing the phoneme identification method of the present invention, FIG. 2 is a diagram showing an example of input audio signal power, FIG. 3 is a diagram showing an example of input audio signal slam, and FIG. 4 is a diagram showing an example of input audio signal power. 5 is a diagram showing an example of a temporal change in a frequency spectrum, FIG. 5 is a diagram showing an example of a time-series pattern of a spectrum, FIG. 6 is a diagram showing an example of a power vector codebook, and FIG. 7 is an example of a time-series feature vector codebook. FIG. 8 is a block diagram showing the configuration of an apparatus for carrying out a conventional phoneme identification method, and FIG. 9 is a diagram showing an example of a phoneme table. In the figure, 101... audio input terminal, 102... own analysis section, 103... power vector calculation section, 104...
- Power vector V Q unit, 105... Power vector codebook, 106... Time series feature vector generation unit, 107... Time series feature vector quantization, 108...
. . . Time series feature vector codebook, 109 . . . Phoneme conversion unit, 110 . . . Phoneme table. Patent applicant Oki Electric Industry Co., Ltd. Agent Patent attorney Takashi Kumagai (1 other person) → fil: L
h 7 good (f spare) + 1/9'l-3'l fJ7S it・%゛Fig. Toru r tochi d'lA: 11 *j Hi T 5 K: 1
Example n Grxv 7'077m Figure 8

Claims (1)

[Scope of Claims] (a) An analysis that frequency-analyzes input audio and calculates a feature vector that is a vector of frequency components of the input audio and power representing the strength of the input audio at fixed time intervals called frames. (b) a process of calculating a power vector obtained by combining the audio powers of adjacent frames; and (c) a process of calculating a time-series feature vector obtained by combining the feature vectors of adjacent frames. , (d) creating an input power pattern and an input time-series feature pattern by performing the processes (a) and (c) on the input voice to be identified; and (e) creating an input power pattern and an input time-series feature pattern on the input voice to be identified. , vector quantizes a large amount of audio data using a power vector codebook created in advance,
(f) vector quantizing the input time series feature pattern using a vector codebook of time series features according to the input power code of the corresponding frame, and obtaining an input time series feature code; and (g) determining the likelihood of a phoneme for each frame from the input power code and the input time-series feature code.