Coupled hidden Markov model for audiovisual speech recognition

Abstract

A speech recognition method includes use of synchronous or asynchronous audio and a video data to enhance speech recognition probabilities. A two stream coupled hidden Markov model is trained and used to identify speech. At least one stream is derived from audio data and a second stream is derived from mouth pattern data. Gestural or other suitable data streams can optionally be combined to reduce speech recognition error rates in noisy environments.

Description

FIELD OF THE INVENTION

The present invention relates to speech recognition systems. More specifically, this invention relates to coupled hidden Markov model techniques for evaluating audiovisual material.

BACKGROUND

Commercial speech recognition systems that primarily use audio input are widely available, but often underutilized because of reliability concerns. Such conventional audio systems are adversely impacted by environmental noise, often requiring acoustically isolated rooms and consistent microphone positioning to reach even minimally acceptable error rates in common speech recognition tasks. The success of the currently available speech recognition systems is accordingly restricted to relatively controlled environments and well defined applications such as dictation or small to medium vocabulary voice-based control commands (hand free dialing, menu navigation, GUI screen control). These limitations have prevented the widespread acceptance of speech recognition systems in acoustically uncontrolled workplace or public sites.

In recent years, it has been shown that the use of visual information together with audio information significantly improve the performance of speech recognition in environments affected by acoustic noise. The use of visual features in conjunction with audio signals takes advantage of the bimodality of the speech (audio is correlated with lip position) and the fact that visual features are invariant to acoustic noise perturbation.

Various approaches to recovering and fusing audio and visual data in audiovisual speech recognition (AVSR) systems are known. One popular approach relies on mouth shape as a key visual data input. Unfortunately, accurate detection of lip contours is often very challenging in conditions of varying illumination or during facial rotations. Alternatively, computationally intensive approaches based on gray scale lip contours modeled through principal component analysis, linear discriminant analysis, two-dimensional DCT, and maximum likelihood transform have been employed to recover suitable visual data for processing.

Fusing the recovered visual data with the audio data is similarly open to various approaches, including feature fusion, model fusion, or decision fusion. In feature fusion, the combined audiovisual feature vectors are obtained by concatenation of the audio and visual features, followed by a dimensionality reduction transform. The resultant observation sequences are then modeled using a hidden Markov model (HMM) technique. In model fusion systems, multistream HMM using assumed state synchronous audio and video sequences is used, although difficulties attributable to lag between visual and audio features can interfere with accurate speech recognition. Decision fusion is a computationally intensive fusion technique that independently models the audio and the visual signals using two HMMs, combining the likelihood of each observation sequence based on the reliability of each modality.

BRIEF DESCRIPTION OF THE DRAWINGS

The inventions will be understood more fully from the detailed description given below and from the accompanying drawings of embodiments of the inventions which, however, should not be taken to limit the inventions to the specific embodiments described, but are for explanation and understanding only.

FIG. 1 generically illustrates a procedure for audiovisual speech recognition;

FIG. 2 illustrates a procedure for visual feature extraction, with diagrams representing feature extraction using a masked, sized and normalized mouth region;

FIG. 3 schematically illustrates an audiovisual coupled HMM; and

FIG. 4 illustrates recognition rate using a coupled HMM model.

DETAILED DESCRIPTION

As seen with respect to the block diagram of FIG. 1, the present invention is a process 10 for audiovisual speech recognition system capable of implementation on a computer based audiovisual recording and processing system 20. The system 20 provides separate or integrated camera and audio systems for audiovisual recording 12 of both facial features and speech of one or more speakers, in real-time or as a recording for later speech processing. Audiovisual information can be recorded and stored in an analog format, or preferentially, can be converted to a suitable digital form, including but not limited to MPEG-2, MPEG-4, JPEG, Motion JPEG, or other sequentially presentable transform coded images commonly used for digital image storage. Low cost, low resolution CCD or CMOS based video camera systems can be used, although video cameras supporting higher frame rates and resolution may be useful for certain applications. Audio data can be acquired by low cost microphone systems, and can be subjected to various audio processing techniques to remove intermittent burst noise, environmental noise, static, sounds recorded outside the normal speech frequency range, or any other non-speech data signal.

In operation, the captured (stored or real-time) audiovisual data is separately subjected to audio processing and visual feature extraction 14. Two or more data streams are integrated using an audiovisual fusion model 16, and training network and speech recognition module 18 are used to yield a desired text data stream reflecting the captured speech. As will be understood, data streams can be processed in near real-time on sufficiently powerful computing systems, processed after a delay or in batch mode, processed on multiple computer systems or parallel processing computers, or processed using any other suitable mechanism available for digital signal processing.

Software implementing suitable procedures, systems and methods can be stored in the memory of a computer system as a set of instructions to be executed. In addition, the instructions to perform procedures described above could alternatively be stored on other forms of machine-readable media, including magnetic and optical disks. For example, the method of the present invention could be stored on machine-readable media, such as magnetic disks or optical disks, which are accessible via a disk drive (or computer-readable medium drive). Further, the instructions can be downloaded into a computing device over a data network in a form of compiled and linked version. Alternatively, the logic could be implemented in additional computer and/or machine readable media, such as discrete hardware components as large-scale integrated circuits (LSI's), application-specific integrated circuits (ASIC's), or firmware such as electrically erasable programmable read-only memory (EEPROM's).

One embodiment of a suitable visual feature extraction procedure is illustrated with respect to FIG. 2. As seen in that Figure, feature extraction 30 includes face detection 32 of the speaker's face (cartoon FIG. 42) in a video sequence. Various face detecting procedures or algorithms are suitable, including pattern matching, shape correlation, optical flow based techniques, hierarchical segmentation, or neural network based techniques. In one particular embodiment, a suitable face detection procedure requires use of a Gaussian mixture model to model the color distribution of the face region. The generated color distinguished face template, along with a background region logarithmic search to deform the template and fit it with the face optimally based on a predetermined target function, can be used to identify single or multiple faces in a visual scene.

After the face is detected, mouth region discrimination 34 is usual, since other areas of the face generally have low or minimal correlation with speech. The lower half of the detected face is a natural choice for the initial estimate of the mouth region (cartoon FIG. 44). Next, linear discriminant analysis (LDA) is used to assign the pixels in the mouth region to the lip and face classes (cartoon FIG. 46). LDA transforms the pixel values from the RGB space into an one-dimensional space that best discriminates between the two classes. The optimal linear discriminant space is computed using a set of manually segmented images of the lip and face regions.

The contour of the lips is obtained through a binary chain encoding method followed by a smoothing operation. The refined position of the mouth corners is obtained by applying the corner finding filter:

$w\left[m,n\right]=\exp \left(-\frac{m^2+n^2}{2{\sigma }^2}\right),{\sigma }^2=70,-3<m,n\le 3,$ in a window around the left and right extremities of the lip contour. The result of the lip contour and mouth corners detection is illustrated in figure cartoon 48 by the dotted line around the lips and mouth.

The lip contour and position of the mouth corners are used to estimate the size and the rotation of the mouth in the image plane. Using the above estimates of the scale and rotation parameters of the mouth, masking, resizing, rotation and normalization 36 is undertaken, with a rotation and size normalized gray scale region of the mouth (typically 64×64 pixels) being obtained from each frame of the video sequence. A masking variable shape window is also applied, since not all the pixels in the mouth region have the same relevance for visual speech recognition, with the most significant information for speech recognition being contained in the pixels inside the lip contour. The masking variable shape window used to multiply the pixels values in the gray scale normalized mouth region is described as:

$\begin{array}{cc}w\left[i,j\right]=\{\begin{array}{c}1,\mathrm{if}i,j\mathrm{are}\mathrm{inside}\mathrm{the}\mathrm{lip}\mathrm{contour},\\ 0,\mathrm{otherwise}\end{array}& \text{(Eq. 1)}\end{array}$ Cartoon FIG. 50 in FIG. 2 illustrates the result of the rotation and size normalization and masking steps.

Next, multiclass linear discriminant analysis 38 is performed on the data. First, the normalized and masked mouth region is decomposed in eight blocks of height 32 pixels and width 16 pixels, and a two dimension discrete cosine transform (2D-DCT) is applied to each of these blocks. A set of four 2D-DCT coefficients from a window of size 2×2 in the lowest frequency in the 2D-DCT domain are extracted from each block. The resulting coefficients extracted are arranged in a vector of size 32. In the final stage of the video features extraction cascade the multi class LDA is applied to the vectors of 2D-DCT coefficients. Typically, the classes of the LDA are associated to words available in the speech database. A set of 15 coefficients, corresponding to the most significant generalized eigenvalues of the LDA decomposition are used as visual observation vectors.

The following table compares the video-only recognition rates for several visual feature techniques and illustrates the improvement obtained by using the masking window and the use of the block 2D-DCT coefficients instead of 1D-DCT coefficients

Video Features            Recognition Rate
1D DCT + LDA              41.66%
Mask, 1D DCT + LDA        45.17%
2D DCT blocks + LDA       45.63%
Mask, 2D DCT blocks + LDA 54.08%

In all the experiments the video observation vectors were modeled using a 5 state, 3 mixture left-to-right HMM with diagonal covariance matrices.

After face detection, processing, and upsampling of data to audio data rates (if necessary), the generated video data must be fused with audio data using a suitable fusion model. In one embodiment, a coupled hidden Markov model (HMM) is useful. The coupled HMM is a generalization of the HMM suitable for a large scale of multimedia applications that integrate two or more streams of data. A coupled HMM can be seen as a collection of HMMs, one for each data stream, where the discrete nodes at time t for each HMM are conditioned by the discrete nodes at time t₁ of all the related HMMs. Diagram 60 in FIG. 3 illustrates a continuous mixture two-stream coupled HMM used in our audiovisual speech recognition system. The squares represent the hidden discrete nodes while the circles describe the continuous observable nodes. The hidden nodes can be conditioned temporally as coupled nodes and to the remaining hidden nodes as mixture nodes. Mathematically, the elements of the coupled HMM are described as:

$\begin{array}{cc}{\pi }_0^c\left(i\right)=P\left(O_0^c|q_t^c=i\right)& \text{(Eq. 2)}\\ b_t^c\left(i\right)=P\left(O_t^c|q_t^c=i\right)& \text{(Eq. 3)}\\ a_i^c{|}_{j,k}=P\left(q_t^c=i|q_{t-1}^0=j,q_{t-1}^1=k\right)& \text{(Eq. 4)}\end{array}$ where

$q_t^c$ is the state of the couple node in the cth stream at time t. In a continuous mixture with t=T−1 t=T−2, . . . t, . . . t=1, t=0, . . . Gaussian components, the probabilities of the coupled nodes are given by:

$\begin{array}{cc}{b}_t^c\left(i\right)=\sum _{m=1}^{M_i^c}{w}_{i,m}^{c}N\left(O_t^c,{\mu }_{i,m}^c,U_{i,m}^c\right)& \text{(Eq. 5)}\end{array}$ where

${\mu }_{i,m}^c$ and

$U_{\underset{\_}{i},\underset{\_}{m}}^c$ are the mean and covariance matrix of the ith state of a coupled node, and mth component of the associated mixture node in the cth channel.

$M_{\hspace{0.17em}\underset{\_}{i}}^c$ is the number of mixtures corresponding to the ith state of a coupled node in the cth stream and the weight

$w_{i,m}^c$ represents the conditional probability

$P\left(p_t^c=m|q_t^c=i\right)$ where

$p_t^c$ is the component of the mixture node in the cth stream at time t.

The constructed HMM must be trained to identify words. Maximum likelihood (ML) training of the dynamic Bayesian networks in general and of the coupled HMMs in particular, is well understood. Any discrete time and space dynamical system governed by a hidden Markov chain emits a sequence of observable outputs with one output (observation) for each state in a trajectory of such states. From the observable sequence of outputs, the most likely dynamical system can be calculated. The result is a model for the underlying process. Alternatively, given a sequence of outputs, the most likely sequence of states can be determined. In speech recognition tasks a database of words, along with separate training set for each word can be generated.

Unfortunately, the iterative maximum likelihood estimation of the parameters only converges to a local optimum, making the choice of the initial parameters of the model a critical issue. An efficient method for the initialization of the ML must be used for good results. One such method is based on the Viterbi algorithm, which determines the optimal sequence of states for the coupled nodes of the audio and video streams that maximizes the observation likelihood. The following steps describe the Viterbi algorithm for the two stream coupled HMM used in one embodiment of the audiovisual fusion model. As will be understood, extension of this method to stream coupled HMM is straightforward.

Initialization

$\begin{array}{cc}{\delta }_0\left(i,j\right)={\pi }_0^a\left(i\right){\pi }_0^v\left(j\right)b_t^a\left(i\right)b_t^v\left(j\right)& \text{(Eq. 6)}\\ {\psi }_0\left(i,j\right)=0& \text{(Eq. 7)}\\ {\delta }_t\left(i,j\right)=\underset{k,l}{\max }\left\{{\delta }_{t-1}\left(k,l\right)a_i{{}_{k,l}{a}_j}_{k,l}\right\}{b}_t^a\left(k\right)b_t^v\left(l\right)& \text{(Eq. 8)}\\ {\psi }_t\left(i,j\right)=\arg \underset{k,l}{\max }\left\{{\delta }_{t-1}\left(k,l\right)a_i{{}_{k,l}{a}_j}_{k,l}\right\}& \text{(Eq. 9)}\end{array}$

Termination

$\begin{array}{cc}P=\underset{i,j}{\max }\left\{{\delta }_T\left(i,j\right)\right\}& \text{(Eq. 10)}\\ \left\{q_T^a,q_T^v\right\}=\arg \underset{i,j}{\max }\left\{{\delta }_T\left(i,j\right)\right\}& \text{(Eq. 11)}\end{array}$

Backtracking (Reconstruction)

$\begin{array}{cc}\left\{q_t^a,q_t^v\right\}={\psi }_{t+1}\left(q_{t+1}^a,q_{t+1}^v\right)& \text{(Eq. 12)}\end{array}$ The segmental K-means algorithm for the coupled HMM proceeds as follows:Step 1—For each training observation sequence r, the data in each stream is uniformly segmented according to the number of states of the coupled nodes. An initial state sequence for the coupled nodes

$Q=q_{r,0}^{a,v},\dots ,q_{r,t}^{a,v},\dots q_{r,T-1}^{a,v}$ is obtained. For each state i of the coupled nodes in stream c the mixture segmentation of the data assigned to it obtained using the K-means algorithm with

$M_i^c$ clusters.Consequently, the sequence of mixture components

$P=p_{0,r}^{\hspace{0.17em}{\hspace{0.17em}}^{-}a,v},\dots ,p_{r,t}^{a,v},\dots p_{r,T-1}^{\stackrel{\_}{a},v}$ for the mixture nodes is obtained.Step 2—The new parameters are estimated from the segmented data:

$\begin{array}{cc}{\mu }_{i,m}^{a,v}=\frac{\sum _{r,t}{\gamma }_{r,t}^{a,v}\left(i,m\right)O_t^{a,v}}{\sum _{r,t}{\gamma }_{r,t}^{a,v}\left(i,m\right)}& \text{(Eq. 13)}\\ {\sigma }_{i,m}^{2^{a,v}}=\frac{\sum _{r,t}{\gamma }_{r,t}^{a,v}\left(i,m\right)\left(O_t^{a,v}-{\mu }_{i,m}^{a,v}\right){\left(O_t^{a,v}-{\mu }_{i,m}^{a,v}\right)}^T}{\sum _{r,t}{\gamma }_{r,t}^{a,v}\left(i,m\right)}& \text{(Eq. 14)}\\ w_{i,m}^{a,v}=\frac{\sum _{r,t}{\gamma }_{r,t}^{a,v}\left(i,m\right)}{\sum _{r,t}\sum _m{\gamma }_{r,t}^{a,v}\left(i,m\right)}& \text{(Eq. 15)}\\ {a_{i}}_{k,l}^{a,v}=\frac{\sum _{r,t}{\epsilon }_{r,t}^{a,v}\left(i,k,l\right)}{\sum _{r,t}\sum _k\sum _l{\epsilon }_{r,t}^{a,v}\left(i,k,l\right)}& \text{(Eq. 16)}\end{array}$ and where

$\begin{array}{cc}{\gamma }_{r,t}^{a,v}\left(i,m\right)=\{\begin{array}{c}1,\mathrm{if}{q}_{r,t}^{a,v}=i,p_{r,t}^{a,v}=m,\\ 0,\mathrm{otherwise}\end{array}& \text{(Eq. 17)}\\ {\epsilon }_{r,t}^{a,v}\left(i,k,l\right)=\{\begin{array}{c}1,\mathrm{if}{q}_{r,t}^{a,v}=i,q_{r,t-1}^a=k,q_{r,t-1}^v=l\\ 0,\mathrm{otherwise}\end{array}& \text{(Eq. 18)}\end{array}$ Step 3—At consecutive iterations an optimal sequence Q of the coupled nodes are obtained using the Viterbi algorithm (which includes Equations 7 through 12). The sequence of mixture component P is obtained by selecting at each moment T the mixture

$\begin{array}{cc}{p}_{r,t}^{a,v}=\underset{m=1,\dots ,M_i^{a,v}}{\max }P\left(O_t^{a,v}|q_{r,t}^{a,v}=i,m\right)& \text{(Eq. 19)}\end{array}$ Step 4—The iterations in steps 2 through 4 inclusive are repeated until the difference between observation probabilities of the training sequences falls below the convergence threshold.

Word recognition is carried out via the computation of the Viterbi algorithm (Equations 7-12) for the parameters of all the word models in the database. The parameters of the coupled HMM corresponding to each word in the database are obtained in the training stage using clean audio signals (SNR=30 db). In the recognition stage the input of the audio and visual streams is weighted based on the relative reliability of the audio and visual features for different levels of the acoustic noise. Formally the state probability at time t for an observation vector

$O_t^{a,v}$ becomes

$\begin{array}{cc}{\tilde{b}}_t^{a,v}\left(i\right)={b_t\left(O_t^{a,v}|q_t^{a,v}=i\right)}^{{\alpha }_{a,v}}& \text{(Eq. 20)}\end{array}$ whereα_a+α_v=1 and α_a, α_v≧0are the exponents of the audio and video streams. The values ofα_a, α_vcorresponding to a specific signal to noise ratio (SNR) are obtained experimentally to maximize the average recognition rate. In one embodiment of the system, audio exponents were optimally found to be

	SNR(db)	30	26	20	16
	αa	0.9	0.8	0.5	0.4

Experimental results for speaker dependent audiovisual word recognition system on 36 words in a database have been determined. Each word in the database is repeated ten times by each of the ten speakers in the database. For each speaker, nine examples of each word were used for training and the remaining example was used for testing. The average audio-only, video-only and audiovisual recognition rates are presented graphically in chart 70 of FIG. 4 and the table below. In chart 70, the triangle data point represents a visual HMM, the diamond data point represents an audio HMM, the star data point represents an audiovisual HMM, and the square shaped data point illustrates an audiovisual coupled HMM.

	SNR(db)	30	26	20	16
	V HMM	53.70%	53.70%	53.70%	53.70%
	A HMM	97.46%	80.58%	50.19%	28.26%
	AV HMM	98.14%	89.34%	72.21%	63.88%
	AV CHMM	98.14%	90.72%	75.00%	69.90%

As can be seen from inspection of the chart 70 and the above table, for audio-only speech recognition the acoustic observation vectors (13 MFCC coefficients extracted from a window of 20 ms) are modeled using a HMM with the same characteristics as the one described for video-only recognition. For the audio-video recognition, a coupled HMM with states for the coupled nodes in both audio and video streams, no back transitions, and three mixture per state, is used. The experimental results indicate that the coupled HMM-based audiovisual speech recognition rate increases by 45% the audio-only speech recognition at SNR of 16 db. Compared to the multistream HMM, the coupled HMM-based audiovisual recognition systems shows consistently better results with the decrease of the SNR reaching a nearly 7% reduction in word error rate at 16 db.

As will be appreciated, accurate audiovisual data to text processing can be used to enable various applications, including provision of robust framework for systems involving human computer interaction and robotics. Accurate speech recognition in high noise environments allows continuous speech recognition under uncontrolled environments, speech command and control devices such as hand free telephones, and other mobile devices. In addition the coupled HMM can be applied to a large number of multimedia applications that involve two or more related data streams such as speech, one or two hand gesture and facial expressions. In contrast to a conventional HMM, the coupled HMM can be readily configured to take advantage of the parallel computing, with separate modeling/training data streams under control of separate processors.

As will be understood, reference in this specification to “an embodiment,” “one embodiment,” “some embodiments,” or “other embodiments” means that a particular feature, structure, or characteristic described in connection with the embodiments is included in at least some embodiments, but not necessarily all embodiments, of the invention. The various appearances “an embodiment,” “one embodiment,” or “some embodiments” are not necessarily all referring to the same embodiments.

If the specification states a component, feature, structure, or characteristic “may”, “might”, or “could” be included, that particular component, feature, structure, or characteristic is not required to be included. If the specification or claim refers to “a” or “an” element, that does not mean there is only one of the element. If the specification or claims refer to “an additional” element, that does not preclude there being more than one of the additional element.

Those skilled in the art having the benefit of this disclosure will appreciate that many other variations from the foregoing description and drawings may be made within the scope of the present invention. Accordingly, it is the following claims, including any amendments thereto, that define the scope of the invention.

Claims

1. A speech recognition method comprising obtaining a first data stream of speech data and a second data stream of face image data while a speaker is speaking;extracting visual features from the second data stream by masking, resizing, rotating, and normalizing a mouth region in a face image, and by using a two-dimensional discrete cosine transform; andapplying a two stream coupled hidden Markov model to the first and second data streams for speech recognition, wherein a Viterbi-based method and a segmental k-means method are used to determine parameters of the two stream coupled hidden Markov model.

2. The method of claim 1, wherein the audio and video data sets providing the first and second data streams are asynchronous.

3. The method of claim 1, further comprising parallel processing of the first and second data streams.

4. The method of claim 1, wherein extracting visual features comprises using linear discriminant analysis.

5. An article comprising a computer readable medium to store computer executable instructions, the instructions defined to cause a computer to obtain a first data stream of speech data and a second data stream of face image data while a speaker is speaking;extract visual features from the second data stream by masking, resizing, rotating, and normalizing a mouth region in a face image, and by using a two-dimensional discrete cosine transform; andapply a two stream coupled hidden Markov model to the first and second data streams for speech recognition, wherein a Viterbi-based method and a segmental k-means method are used to determine parameters of the two stream coupled hidden Markov model.

6. The article comprising a computer readable medium to store computer executable instructions of claim 5, wherein the instructions further cause a computer to process asynchronous first and second data streams.

7. The article comprising a computer readable medium to store computer executable instructions of claim 5, wherein the instructions further cause a computer to process in parallel the first and second data streams.

8. The article comprising a computer readable medium to store computer executable instructions of claim 5, wherein the instructions further cause a computer to provide visual feature extraction from the video data set using linear discriminant analysis.

9. A speech recognition system comprising an audiovisual capture module to capture an audio and a video data set that respectively provide a first data stream of speech data and a second data stream of face image data, the audiovisual capture module extracting visual features from the second data stream by masking, resizing, rotating, and normalizing a mouth region in a face image, and by using a two-dimensional discrete cosine transform; anda speech recognition module to apply a two stream coupled hidden Markov model to the first and second data streams for speech recognition, wherein a Viterbi-based method and a segmental k-means method are used to determine parameters of the two stream coupled hidden Markov model.

10. The speech recognition system of claim 9, further comprising asynchronous audio and video data sets.

11. The speech recognition system of claim 9, further comprising parallel processing of the first and second data streams by the speech recognition module.

12. The speech recognition system of claim 9, wherein the audiovisual capture module is further adapted to extract visual features from the second data stream by using linear discriminant analysis.