JPH04284498A - Method of voice recognition - Google Patents

Abstract

PURPOSE:To attain a high recognition rate by less learning by using a universal code book for unspecified speakers at the time of unspecified speaker rough vocabulary speech recognition. CONSTITUTION:A word/phoneme hidden Markov model(HMM) generated according to speech spectra from many speakers has transition in the order of three states A, B, and C from the left to the right; and the transition probability from the state A to the state B is denoted as tAB and the probability of output of a word/phoneme feature vector yt from the state A to the state B at a point (t) of time is denoted as PAB(y1). Similarly, relatively short learnt speeches are inputted, speaker by speaker, and a word/phoneme ergodic hidden Markov model (speaker HMM) generated according to the speech spectra has transition between two states 1 and 2; and the probability of transition from the state 1 to the state 2 is t12 and the probability of the output of the feature vector yt from the state 1 to the state 2 at a point (t) of time is P12(y1). At this time, the HMM (figure 1) in the product space of those unspecified speaker HMM and speaker HMM is generated and the HMM in the product space is used to recognize the input voice of the speaker.

Description

[Detailed description of the invention]

0001

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition method using a hidden Markov model and applying it to speaker-independent large-vocabulary continuous speech recognition to improve recognition performance.

0002

[Prior art] Hidden Markov model (for example, Seiichi Nakagawa's
“Speech Recognition Using Probabilistic Models” edited by the Institute of Electronics, Information and Communication Engineers (19
In speaker-independent speech recognition according to 88)), a codebook created based on speech spectra from many speakers is often used. This codebook is
It is called the universal codebook. However, as shown in FIG. 4B, the codebook space 11 of a particular speaker is a subspace of the universal codebook 12, as shown in FIG. 4B. Furthermore, the movement of codewords within the codebook 12 is also unique for each speaker.

Despite this fact, in speaker-independent speech recognition using Hidden Markov Models (HMM),
Hidden Markov models of words or phoneme units were created using the Universal Codebook 12 and a large amount of speech data from many speakers. Therefore, the constraints of the speaker-specific codebook space 11 are not considered at all,
It caused various side effects, including deterioration of recognition performance in speaker-independent, large-vocabulary continuous speech recognition.

0004

[Means for Solving the Problems] According to the present invention, a hidden Markov model representing phonemes/words for an unspecified speaker and a hidden Markov model representing characteristics of the speaker are synthesized, and the synthesized model is used. to perform speech recognition of the speaker. In other words,
The basic formula for statistical continuous speech recognition considering the speaker according to the present invention is written as follows. Here, the term regarding S is introduced in this invention.

P(W,S|Y)=P(W,S)P(Y|W,
S)/P(Y)
=P(S)P(W|S)P(Y|W,S)/P(Y
) Here, W: Word string S: Speaker Y: Vector sequence of input speech P(S): Probability that speaker S uses this speech recognition device P(W|S): Speaker S exists. It is considered as the probability of generating the word string W, and is calculated by a statistical language model by the speaker S (for example, Kano, "Speech Recognition Using Statistical Methods", Journal of the Institute of Electronics, Information and Communication Engineers, Vol. 73, No. 12, pp 1276-1285).
, (1990.12)).

P(Y|W,S): Speaker S with occurrence content W
According to the probability (acoustic model) of the vector sequence Y of input speech at
(S)P(W|S)P(Y|W,S)}W,S The word string W is estimated using the information S of the speaker.

[0007] Here, P(S) represents the probability that the speaker S is using this speech recognition device. Furthermore, the acoustic model (word/phonological model) P(Y|W,S) by speaker S
Consider modeling using a hidden Markov model (HMM). If a large amount of word/phoneme speech data is generated for each speaker, it is possible to create a word/phoneme HMM for each speaker, but normally it is not possible to generate a large amount of speech data for all speakers. is not realistic. Therefore,
As is commonly done, we use a word/phoneme HMMP (Y|W) created using speech data from a large number of speakers, and use a speaker-specific codebook for this P(Y|W). Think about limiting your space. Hereinafter, we will consider using the HMM to represent the space of a speaker-specific codebook and the movement of codewords.

[0008] The HMM is represented by the following six sets. HMM: M=(U, V, T, P, I, F) Here, U:
Set of states V: Set of input vectors T: Set of transition probabilities P: Set of output probabilities I: Initial state F: Final state Further, the input sequence is expressed as Y=y1 y2 ...yt ...yN.

[0009] Here, as an HMM representing speaker characteristics, we will consider an ergodic HMM whose parameters are estimated from arbitrary utterances for each speaker. A simple example of this ergodic HMM is shown in FIG. 3A. In other words, input a relatively short training speech for each speaker, and let t12 be the transition probability from state 1 to state 2, and let P12(yt ) be the probability that the input vector yt at time t1 will transition from state 1 to state 2. , a model (ergodic HMM) that transitions between states 1, 2, and 3 is created for each speaker. This speaker HMM is defined as Msi = (Usi, V, Tsi, Psi, I
si, Fsi): Represented by speaker i (i=1,..., L). As a word/phoneme HMM estimated from the parameters of speech data of a large number of speakers, consider an HMM as shown in FIG. 3B with a left-to-right transition. This is created from voice data from multiple speakers. This word/phonetic H
Let MM be Mpj = (Upj, V, Tpj, Ppj,
Ipj, Fpj): word/phoneme j (j=1,...,
M).

[0010] In this invention, these two HMMs, Msi
Create an HMM in the product space of and Mpj. This synthesis H
The MM will be referred to as the speaker-constrained word/phoneme HMM, and is defined as follows. Mpji=(Upji, V, Tpji, Ppji, Ip
ji, Fpji): word/phoneme (j), speaker (i)
=(Upj×Usi,V,Tpj×Tsi,Λ(Ppj
× Psi), Ipj × Isi, Fpj × Fsi) Then, the speech uttered by speaker i can be recognized by finding the speaker i and word/phoneme j that maximize the probability using this speaker constraint word/phoneme HMM. conduct. In the above equation, Λ() is a scale factor that makes the sum of output probabilities equal to 1.

[0011]

[Embodiment] In this invention, as described above, an HMM is used which is a combination of an HMM representing phonemes/words for an unspecified speaker and an ergodic speaker characteristic HMM representing characteristics of the speaker.
These were synthesized using an ergodic speaker-specific HMM consisting of two states 1 and 2 as shown in A, and a word/phoneme HMM for non-specific speakers consisting of three states A, B, and C shown in Fig. 2B. Figure 1 shows an example of the configuration of a speaker-constrained word/phoneme HMM with six states.
Shown below. The calculation formulas for the values of transition probability and output probability are shown in the figure. However, although transition probabilities and output probabilities are omitted for some transitions, they can be calculated in the same way.

An example of a speech recognition device using such a speaker-constrained word/phoneme HMM is shown in FIG. 4A. The audio input from the input terminal 1 is converted into a digital signal in the feature extraction unit 2, subjected to LPC cepstrum analysis, and then vector quantized using a universal codebook for each frame (10 milliseconds). speaker model HM
The learning unit 3 of M selects the speaker HMM with the highest likelihood from a plurality of ergodic speaker HMMs stored in advance, and also performs additional learning on that HMM based on input speech. Next, in the speaker-constrained phoneme HMM synthesis unit 4, the speaker model HMM and the speaker-independent phoneme model HM are combined according to the present invention.
A speaker-constrained phoneme HMM is synthesized from M5. The continuous speech recognition unit 6 uses this speaker-constrained phoneme HMM to recognize the content of the input speech, and outputs a recognition result 7.

[0013] In this embodiment, a procedure was shown in which continuous speech recognition is performed after synthesizing speaker-constrained phonemes HMM in the synthesizing section 4. However, during continuous speech recognition, speaker-constrained phonemes HM
It is also possible to include a procedure for synthetically creating M.

[0014]

[Effects of the Invention] As described above, according to the present invention,
By using speaker HMM, phonology/
It is possible to constrain the word HMM to the spatial and spectral movements specific to the generator, and it is possible to achieve a high recognition rate. This technique eliminates the need for a speaker to generate large amounts of speech data in order to create a phoneme/word HMM for a particular speaker. According to this method, a speaker model is selected using a small amount of arbitrary speech data, the speaker model is adapted through additional learning, and this speaker HMM model is synthesized with a phoneme/word HMM. It becomes possible to create a highly accurate phoneme/word HMM for a specific speaker.

[0015] Although the above explanation mainly focused on discrete HMM, this method is based on fuzzy vector quantization based HMM.
It can be similarly applied to MM and continuous distribution HMM. Similarly, the present invention can be applied to general speech recognition methods that satisfy the conditions constrained by two HMMs, and can be used, for example, to adapt to noisy environments, types of microphones, etc., and improve recognition performance. In other words, recognition performance can be improved by using, for example, an ergodic HMM representing characteristics of a microphone instead of an ergodic HMM representing characteristics of a speaker. In addition to speech recognition, the present invention can also be applied to objects for which constraint conditions are given by two HMMs.

[Brief explanation of the drawing]

FIG. 1 is a diagram showing a configuration example of a speaker-constrained word/phoneme HMM that is a combination of a word/phoneme HMM of an uncharacteristic speaker and an ergodic speaker-specific HMM.

FIG. 2A is a diagram showing an example of the configuration of an ergodic speaker-specific HMM, and FIG. 2B is a diagram showing an example of the configuration of a word/phoneme HMM for an unspecified speaker.

FIG. 3A is a diagram showing a simple example of an ergodic HMM representing speaker characteristics, and FIG. 3B is a diagram showing a simple example of a word/phoneme HMM.

FIG. 4A is a block diagram showing an example of a continuous speech recognition system to which the present invention is applied, and FIG. 4B is a diagram showing an example of the relationship between a universal codebook and a speaker-specific codebook space.

Claims (1)

[Claims] Claim 1: A speaker-independent speech recognition method comprising: a hidden Markov model representing phonemes/words for speaker-independent;
A speech recognition method characterized by synthesizing a hidden Markov model representing characteristics of a speaker and performing speech recognition of the speaker using the synthesized model.