JP3247746B2 - Creating a noise-resistant phoneme model - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a technique for creating a phoneme model used for speech recognition, and more particularly, to a technique for creating a noise-resistant phoneme model effective for speech recognition in a place where noise exists.

[0002]

2. Description of the Related Art A conventional speech recognition system using a phoneme HMM will be described with reference to the drawings. FIG. 6 is a diagram showing an example of a functional block diagram for realizing a conventional voice recognition system. 1 in the figure
Is a voice input unit, 2 is a recognition unit, and 3 is a recognition result output unit. Reference numeral 4 denotes a phoneme model storage unit in which the phoneme HMM is stored.

First, the operation of FIG. 5 will be described. In the recognition unit 2, the speech input to the speech input unit 1 is framed, and for each frame, a probability calculation is performed using a phoneme HMM stored in advance in the phoneme model storage unit 4, and a speech with a high probability is used as a recognition result. It has been transmitted to the recognition result output unit 3. The operation of the recognition unit 2 will be described in more detail. When a voice is input for a certain period, the input is divided into unit times to form a frame, and a feature parameter is extracted for each frame. Then, for the feature parameters extracted for the first frame, for example, using a phoneme HMM prepared for each phoneme, the probability of “A”, the probability of “I”,
The probability of "U" is obtained for each phoneme.

Subsequently, the characteristic parameters of the next frame are input, and the probability that this frame will come after the previous state is obtained for each phoneme. This is performed for all the frames of the input voice. For example, if the probability is highest when using the phonetic HMM for the phoneme “a”, the input voice is recognized as “a”.

[0005]

However, conventionally, the phoneme HM
M is created based on voice information obtained in a noise-free state. In a real environment, the performance of voice recognition is significantly reduced due to the influence of noise. Also, in the method of recording speech under various kinds of noise in advance and generating a phonological HMM from the speech, since the type of noise is enormous, the system becomes bloated in order to obtain high recognition performance. is there.

[0006] The present invention solves the above-mentioned problems and can be used in a simple system, thereby creating a phoneme model capable of obtaining high recognition performance even in a noisy situation, thereby producing a vocal environment. The purpose of this invention is to create a phoneme model in accordance with, and to improve speech recognition performance.

[0007]

In order to achieve the above object, a method for creating a noise-resistant phoneme hidden Markov model according to the present invention employs a pre-created speech utterance in a noise-free environment.
Phonetic hidden Markov model and hidden Markov model of noise
The state and the state of the phonetic hidden Markov model of the speech
Combine and combine states of hidden Markov model with noise
In the spectral domain for the probability of each state
Combining in the product space by adding probabilities,
Create a folded phoneme hidden Markov model. Or,
Speech sound in a noise-free environment represented by the pustrum domain
Rhythmic Hidden Markov Model and Hidden Markov Model of Noise
Cosine transform and exponential transform of force probability distribution
To calculate each distribution in the linear spectral domain
And the phonological hidden Markov model of the speech and the noise
The distribution of the Markov model in the linear spectral region.
Construct a phonemic hidden Markov model that is convolved and adapted to noise
And the inverse of the logarithmic transformation
Perform cosine transform to convert to cepstrum domain.
Or, in synthesis, the phonetic hidden Markov model of the speech
Convolve the output probability of the hidden Markov model of the noise
Construct a phonetic hidden Markov model adapted to the noise,
Logarithmic transformation and inverse cosine transformation of distribution obtained by convolution
Conversion to the cepstrum domain.

[0008]

In this way, the phoneme H considering the noise at the utterance location is obtained.
An MM can be created, and the speech recognition performance under noise can be improved.

[0009]

FIG. 1 is a diagram showing a procedure for creating a noise-resistant phoneme HMM according to the present invention. Hereinafter, the creation procedure will be described with reference to FIG. Cepstrum coefficients are widely used as acoustic parameters used in HMMs for speech recognition.
The cepstrum coefficient has a relationship between a logarithmic spectrum (logarithmic power spectrum) and a cosine transform. On the other hand, ambient noise and the like are additive noise and can be added to speech in the spectral domain.

First, in the cepstrum domain, a phonemic HMM in a noise-free environment and a noise HMM in a noisy environment are created, and each of them is cosine-transformed to calculate each logarithmic spectrum. Next, they are subjected to exponential transformation to calculate their respective linear spectra, the output probabilities of the phoneme and noise are obtained by convolution, and the phoneme HM with noise configured in this way is obtained.
By performing logarithmic transformation of M and then performing inverse cosine transformation, a phoneme HMM is created.

These conversions are performed for each of speech and noise, and both parameters are added in this linear spectrum region. Thereafter, the linear spectrum is logarithmically transformed back to a logarithmic spectrum region, and further returned to a cepstrum region by cosine transformation to create a phonemic HMM taking ambient noise into account.

Here, the phonological HMM and the noise HMM are HM using a mixture of so-called Gaussian mixtures.
Let it be represented by M. Therefore, in the above conversion, the conversion is performed for the covariance as well as the conversion of the average value. Next, a specific calculation method will be described. First, these H
Cepstrum coefficients from the 0th to the pth order are considered as parameters of the MM. This vector is represented as C = (C ₀ C ₁ C _{2 -----} C _p-1 C _p ) D = (D ₀ D ₁ D _{2 -----} D _p-1 D _p ). Here, C is a phoneme parameter, and D is a noise parameter. The cosine transform from the cepstrum coefficients to the logarithmic spectrum is known as a cosine transform, which is a linear transform and is represented by a (p + 1) × m transform matrix (COS).
The vectors of the logarithmic spectrum are represented by LC and LD, respectively. Therefore, the average value obtained by the cosine transform is as follows: LC = C (COS) LD = D (COS) Also, the covariances of the cepstrum coefficients are Σ ^C and ^D
Then, the covariances Σ ^LC , Σ ^LD of the logarithmic spectrum are Σ ^LC = (COS) OS ^C (COS) ^{t t LD} = (COS) Σ ^D (COS) ^t . As described above, by converting the average value and the covariance, the average value and the covariance in the logarithmic spectrum region of the normal distribution of speech and noise can be obtained.

Next, an exponential conversion for converting a logarithmic spectrum into a linear spectrum will be described. This transformation does not maintain the shape of the normal distribution, but approximates it with a normal distribution. When the mean SC, SD and the covariance Σ ^SC , Σ ^SD of the distribution form at the time of the exponential transformation are calculated, SC ⁱ = exp (LC ⁱ + Σ ^LC _ii / 2) SD ⁱ = exp (LD ⁱ + Σ ^LD _{ii) ^{/ 2) Σ SC ij = SC}} i × SC j × {exp (Σ LC ij) -1} Σ SD ij = SD i × SD j × {exp (Σ LD ij) -1} becomes. As described above, the sum of speech and additive noise is obtained in the linear spectrum region, and since the sum is a distributed convolution, the average of the results of the two distributed convolutions is easily obtained. Value, M and covariance, Σ ^M
Can be requested. Therefore, it is ^{M = SC i + SD i Σ} M = Σ SC + Σ SD. The mean value and covariance of the distribution form obtained in this way are converted to the cepstrum domain in the reverse of the previous process.

First, logarithmic transformation, which is the inverse transformation of exponential transformation, is performed. The log-transformed average is LM,
If the covariance is ^{LM LM} , it is the inverse of the exponential transform, so LM _i = log (M _i ) −1/2 log (Σ ^M _ii / M _i ² +1) Σ ^LM _ij = log (Σ ^M _ij / (M _i M _j ) +1). Furthermore, the inverse cosine transform (same as cosine transform) (COS ') by mx (p + 1) converts the logarithmic spectrum into a cepstrum region, the average value, the S and covariance ^{Σ S, S = LM (COS} ') Σ ^S = (COS ') Σ ^LM (COS') ^t . Then, the H obtained in the cepstrum region
The normal distribution of the output probability of the MM is taken up to the linear spectrum region, added to the noise model, and then subjected to an inverse transform to produce a phoneme HMM to which the noise is added.

In the above description, two distribution types are taken up,
The synthesis method was described. Next, a method of combining two HMMs will be described. Normally, a phonological HMM is represented by a model of about three states from left to right as shown in FIG. On the other hand, an ergodic HMM as shown in FIG. 3 is suitable as a noise model. Using these two models as a product, consider combining in a product space. Then, an HMM in a product space as shown in FIG. 4 is obtained. Each state is a phoneme H
It consists of a combination of the state of the MM and the state of the noise HMM. Therefore, the output probability of each corresponding state is converted into the above-mentioned linear spectral domain, followed by convolution and inverse conversion.

When the output probability is a single normal distribution, the above-described conversion may be performed for two distributions. When the output distribution is a mixture of normal distributions, the above-described conversion is performed for every pair of distribution combinations. In the description so far, only the cepstrum coefficient has been described. However, the same conversion is applied to the Δcepstrum and Δpower, which are often used in the phonological HMM, since they are represented by the linear sum of the cepstrum coefficients. it can.

FIG. 5 is a diagram showing an embodiment in which the present invention is applied to a speech recognition device. In the figure, 5 is a phoneme model storage unit, 6 is a noise input unit, 7 is a noise-resistant phoneme model creation unit that executes the noise-resistant phoneme model creation method of the present invention, and 8 is a noise-resistant phoneme model storage unit. A phoneme HMM model created in advance from speech information obtained in a noise-free state is prepared in the phoneme model storage unit 5, and the noise-resistant phoneme model creation unit 7 converts the noise phoneme HMM from the noise input from the noise input unit 6. The noise-resistant phoneme model created according to the method of the present invention is stored in the noise-resistant phoneme model storage unit 8.

The recognizing unit 2 performs speech recognition of the input speech input from the speech input unit 1 using the phoneme model stored in the noise-resistant phoneme model storage unit 8 and outputs the speech from the recognition result output unit 3.

[0019]

According to the present invention, since a noise model of the speech environment can be used, a robust phoneme HMM suitable for the speech environment can be created, and a high phoneme recognition rate can be expected to be achieved.

[Brief description of the drawings]

FIG. 1 is a diagram showing a procedure for creating a noise-resistant phoneme model according to the present invention.

FIG. 2 shows an example of a three-state three-loop phoneme HMM.

FIG. 3 shows an example of a noise HMM using a two-state ergodic HMM.

FIG. 4 shows a phoneme HMM obtained by combining the phoneme HMM of FIG. 2 and the noise HMM of FIG. 3 in a product space.

FIG. 5 is a diagram illustrating an embodiment when the present invention is applied to a speech recognition device.

FIG. 6 is a diagram illustrating a configuration of a conventional voice recognition device.

────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Frank Martin 4-6-29 Komaba, Meguro-ku, Tokyo (56) References JP-A-2-83596 (JP, A) Proc IEEE Int Conf Acoustics, Speed , And Signal Process (ICASSP) 1992, Vol. 1 I-489 to I-492 "ADAPTATION ION OF THE HMM DIS TRIBUTIOS: APPLICAT ION TO A VQ CODEBOOK AND TO A NOISY ENVIRONMENT" AU .Cl. ⁷ , DB name) G10L 15/14 G10L 15/06 G10L 15/20

Claims (3)

(57) [Claims] 1. A state of a feature parameter of a speech is obtained for each phoneme.
A creating method of phoneme hidden Markov model consisting of that probability, hidden phoneme of the voice in the absence of noise, which is prepared in advance environment
About the Markov model and the hidden Markov model of noise
Te, of the phoneme hidden Markov model of the speech state of the noise
Combine the states of the Hidden Markov Model and the corresponding
A noise-resistant phoneme model creation method characterized by creating a phoneme hidden Markov model on which noise is superimposed by adding the probabilities in the spectral domain to the probabilities of the respective states to synthesize in a product space . 2. Obtaining a state of a feature parameter of a speech for each phoneme.
Is a method for creating a phonetic hidden Markov model consisting of probabilities, and the cosine of the output probability distributions of the phonetic hidden Markov model and the noise hidden Markov model in a noise-free environment represented by the cepstrum domain. Calculate distribution in linear spectral domain by performing transformation and exponential transformation
And the hidden phonetic Markov model of the speech and the hidden
Convolve the distribution of the Lkov model in the linear spectral domain
A noise-resistant phoneme model creation method, comprising: constructing a phoneme hidden Markov model adapted to noise and performing logarithmic transformation and inverse cosine transformation on the distribution obtained by the convolution to cepstrum domain. 3. The method according to claim 2 , wherein said synthesizing comprises a phonemic hidden Markov of said speech.
Tatami the output probability of the hidden Markov model of the model and the noise
Construct phoneme hidden Markov model adapted to noise
Then, the distribution obtained by the convolution is log-transformed and inversely transformed.
Conversion to the cepstrum domain.
3. A method for creating a noise-resistant phoneme model according to claim 2.