JPH04122995A - Phoneme discriminating method - Google Patents

Abstract

PURPOSE:To obtain high phoneme discrimination performance even by a recognition system for unspecified persons which does not limit a speaker by accurately obtaining characteristic spectrum structure prescribed by a voice characteristic to the speaker. CONSTITUTION:Specially, an optimum feature code selection part 108 is provided, the total feature VQ distortion from the frame at the head of a voice to the frame at the tail is calculated, and the number of a code book where the total distortion is minimum is found. Then a phoneme conversion part 109 converts a power code, optimum code length, and a feature code into a phoneme symbol. Thus, the input voice is converted into a phoneme symbol sequence. Namely, plural spectrum feature code books which are divided by sound quality previously are provided and matched with the input voice quality to perform detailed discrimination and at the same time, a classification of the sound quality is decided, so the spectrum structure of any voice quality can accurately be discriminated.

Description

[Detailed description of the invention] [Industrial application field] The present invention relates to phoneme identification processing.

[Prior art]

A conventional technique of this type was disclosed in ``Evaluation of frame-by-frame phonetic information based on mutual information'', Institute of Electronics, Information and Communication Engineers, Speech Study Group, 5P-103 (1988).

One of the methods that is currently being actively researched in the field of speech recognition is a recognition method that uses phoneme identification processing. The phoneme identification process converts the 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 word searches and sentence types to be recognized.

FIG. 9 is a block diagram showing the configuration of a device that implements 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 79 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 Si, 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.

S j ” (Sil Si2..., 5ij-1
, 5ij) where 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, let O be the frame number at the beginning of the audio, and I be the frame number at the end of the audio (i=...-2, -1, 0, 1, 2,..., I,
I+1. I+2. ...).

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

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

The power vector 79 unit 204 refers to the power vector codebook 205 to obtain vector quantization (VQ) L and power code Ci.

Ci = argmin d (Pi, y-) where,
d (p i, y") is the vector Pi and the vector "
Represents distance.

The power vector quantization process roughly classifies the input audio according to the power shape, and the major classification result is the power food Cf.
becomes. No. 1 is a power food vector, and m is a power code number.

y"!(y"-n+y-n+1. '・1y"-C7"
n+7-L + +y-n-1. y-n) (m
=1.2. -, M) where M is the power vector codebook size.

The feature vector VQ unit 206 uses the feature vector codebook 20
Vector quantization of feature vector i is performed based on 7. Further, 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 is calculated by the following equation.

Z i = armin d (5i, χ(c)'
) (2) Here, (C)' is the power code C
The corresponding shank feature code vector, r is the feature vector code.

X(C)'= (X(c)'+, X(c)'t, -, X
(c)'r-.

X(c)'h ) (r=1.2....,
R(c)) R(c) is the feature vector code size corresponding to power code C.

The phoneme conversion unit 209 converts the phoneme symbol L from these Ci and Zj.
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. Power code C1=1, feature code Z
According to this table, when i=1, the phonetic symbol Li is a. Also, C1=2. In Zisu3, it can be converted as 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 various features of speech are effectively captured and discriminated. When humans hear voices, they not only use the static information of the voice, that is, the strength of the sound at a certain moment, the difference in timbre (spectrum), but also the static information of the voice.
Sound quality classification is determined through a series of sounds. Naturally, a series of utterances has a unique spectral structure defined by the sound quality unique to the speaker. This spectral structure differs depending on the speaker. Therefore, in a recognition system for an unspecified person that does not limit the speaker, it is necessary to accurately capture this spectral structure.

However, the conventional phoneme identification method described above does not take this point into consideration at all. There was only one codebook for capturing spectral features, and any utterance was code-coded based on that single codebook. As a result, combinations of spectral code sequences that would normally not occur in the utterances of a human speaker are often assigned, which is one of the reasons why recognition performance is not improved.

The present invention has been made in view of the above points, and by accurately capturing the unique spectral structure that can be defined by the voice unique to the speaker, it is possible to achieve high phoneme discrimination performance even in a recognition system for an unspecified person without limiting the speaker. It's about getting.

[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 input audio and calculation of 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) Adjacent The process of calculating a power vector obtained by combining the audio powers of the frames, and (c) the processes of (a) and (b) for the input audio to be identified, calculate the input Hower pattern and the input feature pattern. (d) processing for vector quantizing the input power pattern using the power vector codebook to obtain an input power code; (e) processing for generating a plurality of feature vectors for the input feature pattern; a process of vector quantizing using a codebook to obtain a plurality of input feature codes and a plurality of input feature VQ distortions; (f) a feature vector codebook that minimizes the sum of each of the plurality of input feature VQ distortions; and selecting an optimal input feature code according to the codebook; <g) calculating the likelihood of a phoneme for each frame from the optimal input feature hood and the input power code.

[Effect]

According to the present invention, a plurality of spectral feature codebooks are preliminarily divided by voice, and the input voice is checked against the plurality of codebooks to perform detailed identification and at the same time to determine the type of voice. Therefore, it is possible to accurately capture the spectral structure of any sound quality and identify it.

Furthermore, in the present invention, when performing phoneme identification using multiple feature quantities, instead of preparing multiple codebooks for all feature quantities, the present invention uses feature quantities that are less speaker-dependent, such as voice power. Variable features and the like are handled by a single codebook, while features that are strongly speaker-dependent, such as spectral features, are handled by multiple codebooks. In this way, by preparing an appropriate number of codebooks according to the speaker dependence of the feature values, it is possible not only to reduce the amount of processing, but also to make the system more stable because it does not process unnecessary dictionary items. Reliable classification can be performed. Therefore, efficient and higher phoneme discrimination performance can be obtained.

〔Example〕

Embodiments of the present invention will be described below based on 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 figure, the characteristic vector VQ
107 is a feature vector codebook, 108 is an optimal feature code selection section, 109 is a phoneme conversion section, and 110 is a phoneme table.

FIG. 2 is a diagram showing an example of input audio signal power, FIG. 3 is a diagram showing an example of input power pattern, FIG. 4 is a diagram showing an example of an input codebook, and FIG. 7 is a diagram showing an example of a feature vector codebook. It is.

An audio signal that can be inputted from the audio input terminal 101 in FIG. 1 is converted into a feature vector representing characteristics in the analysis section 102. This feature vector derivation method includes a method using multiple bandpass groups with slightly different center frequencies,
Although a method using spectrum analysis using FFT (Fast Founoe Transform) is conceivable, a method using a group of band-pass filters will be exemplified here. Feature vector Si
The in-band frequency components extracted by a group of bandpass filters with different center frequencies are logarithmically transformed,
The data taken out at small intervals called frames is used.

S i = (Sit, Si2.-.5ij-1,5i
j) 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 =J−'ΣSij (
3)” = For simplicity, we set the frame number at the beginning of the audio to 0,
The frame number at the end of the audio is set to 1. (i: ・・−
2, -1, 0, 1, 2,...1. I+1. I+2.・
) The power vector calculation unit 103 combines the powers Pi of adjacent frames to calculate a power vector Pi.

p i = (Pi-n, Pi-n+1.=・, Pi-
1, Pi, Pi+1. -, Pi+n-1, Pi+n), vector quantization (VQ) L and power code Ci are determined by referring to the codebook 105.

Ci = argim d (p i , y”)
(4) Here, d(P i , y"") represents the distance between the vector Pi and the vector ya&. 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 the power code vector, m is the power food number (m=1, 2
．． ..., M) where M is the power vector codebook size.

In the feature vector quantization 106, the feature vector 5i is vector quantized based on the feature hector codebook 107 as shown in FIG. The feature vector codebook 107 has Q types of codebooks created in advance from a large number of voice data grouped according to the characteristics of the speaker, ie, gender, age, sound quality, etc. Furthermore, each codebook further includes sub-codebooks for each power code. In other words, there are Q codebooks in the feature vector codebook 107,
Each codebook is made up of M sub-codebooks. M is the number of power codes mentioned above.

Feature vector quantization processing performs detailed classification of speech based on detailed features, contrary to the general classification of power vector quantization processing. First, Q sub-codebooks corresponding to the power code Ci are selected from the feature vector codebook 107 from each codebook. In other words, this corresponds to switching the detailed identification dictionary to an appropriate one in consideration of the results of the major classification. Quantization (VQ) is performed using the Q selected subcodebooks.

Now, if c=Ci, feature code Z(q)i, feature V
Q strain V(q)i is expressed by the following equation.

Z(q)i=argmin d(S i
, Me (q, c)') (q=l, 2....,
Q) (5)V(q)i=! flin d (,5
' i , X (q, c)' ) (q=1, 2...
, Q) (6) Here, X(q, c)' is the feature code vector of the g-th sub-codebook for power code C,
r is the feature hector code.

me(q, c)' = (x (q, c)' l+ x
(q, c)' 2. −, x (q, c)' +−+−
x (q, c)' + ) (r=1.2..., R(
q, c)) R(Q, C) is the sub-codebook size corresponding to power code C.

The optimal feature code selection unit 10B calculates the sum Wq of the feature vQ distortion V(q)i from the voice start frame to the voice end frame.

Wq=ΣV(q)i (q=1.2.-, Q)
(7);O Furthermore, find the codebook number qm for which Wq is the minimum.

q m=m1n Wq (q=1
．． 2.・=, Q) (8) Quantization (v
Q) The hood has the characteristic code Zi. That is, Z i = V(qm)i
(9) The phoneme conversion unit 109 converts the power code C11 optimum code length qm and feature code Zi into phoneme symbol Li. Various methods can be considered for this process, but the simplest table pull-up method will be explained here. In the phoneme table 110 shown in FIG. 8, when the power food C1-1, the optimal codebook qm=1, and the feature code Zi=1, the phoneme symbol Li is a. Furthermore, when the power code C1=2, the optimal codebook q m =2, and the feature code Zi=2, the phonetic symbol Li becomes e.

In this way, the input speech can be 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) For voice data uttered in advance by many speakers,
Speaker characteristics such as gender, age, sound quality, etc. are degrouped and a speaker group number is assigned.

(b) Calculate a power vector and a feature vector for the audio data, and convert each into power vector data (third
(see figure) and stored as feature data (see figure 5).

(c) clustering the power vector data;
Create a power vector codebook.

(d) The power vector data is vector quantized using a power vector codebook to create /warcode data.

(e) The feature data is divided by the speaker group number and separated as classified feature data.

(f) For each of the classification feature data, the data is further divided according to the power code, and clustering processing is performed on each of the data to create a sub-codebook.

A feature vector codebook 107 is a collection of these subs. (See Figure 7) (g) Vector quantize each classification feature data using a sub-codebook determined from the corresponding power food to obtain feature code data.

(h) A phonological table 1 representing the correspondence between each code based on the phonological code data given to the audio data in advance through inspection etc., the feature code data and the power code data.
Create 10.

〔Effect of the invention〕

As explained above, according to the present invention, the following excellent effects can be obtained.

(1) It has multiple spectral feature codebooks that are pre-divided by sound quality, and compares the input sound quality with those codebooks to perform detailed identification and at the same time to determine the type of sound quality. Even for high-quality voices, it is possible to accurately capture and identify the spectral structure of the voice quality.

(2) Furthermore, when performing phoneme identification using multiple features, instead of preparing multiple codebooks for all features, multiple features with less speaker dependence, such as spectral features, can be used. We are trying to deal with this problem using the codebook. In this way, by preparing an appropriate number of codebooks according to the speaker dependence of the feature values, it is possible to not only reduce the amount of processing, but also improve stability by not processing unnecessary dictionary items. This allows for reliable classification. Therefore, efficient and high phoneme discrimination performance can be obtained.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing the configuration of a device 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 an input power pattern, and FIG. 4 is a diagram showing an example of input power pattern. 5 is a diagram showing an example of a frequency spectrum of a voice, FIG. 5 is a diagram showing an example of a spectrum pattern, FIG. 6 is a diagram showing an example of a power vector codebook, FIG. 7 is a diagram showing an example of a feature vector codebook,
FIG. 8 is a diagram showing an example of a phoneme table, FIG. 9 is a block diagram showing the configuration of an apparatus for implementing a conventional phoneme identification method, and FIG. 10 is a diagram showing an example of the phoneme table. In the figure, 101...Audio input terminal, 102...Analysis section 0, 103...Power vector calculation section, 104...
...Power vector calculation unit, 105...Power vector calculation unit, 106...Feature vector VQ unit, 107
... Feature vector codebook, 108 ... Optimal feature code selection section, 109 ... Phoneme conversion section, 1109.
...Phonological table. Patent Applicant: Oki Electric Industry Co., Ltd. Agent, Patent Attorney: Takashi Kumagai (1 other person) Fig. 4 Subeft) 1v-”

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) processing to calculate a power vector obtained by combining the audio powers of adjacent frames; and (c) processing the input audio to be identified by performing the processing in (a) and (b) above. a process of creating a power pattern and an input feature pattern; (d) a process of vector quantizing the input power pattern using the power vector codebook to obtain an input power code; and (e) a process of calculating the input power code. , vector quantization using a plurality of feature vector codebooks to obtain a plurality of input feature codes and a plurality of input feature quantization (VQ) distortions; ) Selecting a feature vector codebook that minimizes the sum of each with respect to distortion, and selecting an optimal input feature code using the codebook; (g) Processing for each frame from the optimal input feature code and the input power code A phoneme identification method characterized by finding the likelihood of a phoneme.