Visual Speech Recognition: a solution from feature extraction to words classification

Visual Speech Recognition: a solution from feature extraction to words classification LUCIANA GONÇALVES DA SILVEIRA 1, JACQUES FACON 2, DÍBIO LEANDRO BORGES 2 1 FACULDADE CAMBURY Cambury College, Goiânia, Go, Brazil lugon@cultura.com.br 2 PUCPR- Pontifical Catholic University of Parana, Laboratory for Vision and Image Science, CCET PPGIA, Rua Imaculada Conceição 1155 Prado Velho, Curitiba, Pr, Brazil {facon,dibio}@ppgia.pucpr.br Abstract. Audio-visual Speech Recognition has been an active area of research lately. A bit, and yet unsolved, part of this problem is the visual only recognition, or lip reading. Considering an image sequence of a person pronouncing a word, a full image analysis solution would have to segment the mouth area, extract relevant features, and use them to be able to classify the word from those visual features. In this paper we approach this problem by proposing a segmentation technique for the lips contours together with a set of features based on the extracted contours which is able to perform lip reading with promising results. We have collected visual speech sequences in our lab and show the results here for a set of ten words in Brazilian Portuguese, spoken by different speakers in more than 150 samples. The approach can be extended and applied to other spoken languages as well. 1. Introduction Audio-visual speech recognition is an area which embraces the tasks of lip reading and audition. It is a psychological finding, see Summerfield in [12], that automatic speech recognition might be made more robust if visual speech information could be incorporated. However, straight recognition of fluent speech based solely on visual information is limited, since there are not enough distinctions for classification purposes on lips movements whenever one is speaking using a regular size vocabulary. Because of larger availability of commercial dictation systems, and multimedia interfaces with visual input capacity, the problem of designing visual speech recognition solutions has received great research interest lately. For a recent review of the area see Chen in [3] Providing a computational solution to visual speech recognition can be divided into three main tasks, or stages. First, given a sequence of images the mouth area has to be detected automatically; second, features suitable for the identification and classification of the speech ought to be extracted from the images; and finally, recognition is to be achieved based on the visual features regarding a particular vocabulary. In this paper we report a novel solution for visual speech recognition including the three tasks mentioned. The solution was tested for a set of words from Brazilian Portuguese, however it can be extended to work for different sets of words and other languages as well. We review here some important results found in the recent literature most related to the problem and the approach proposed in this paper. Caplier in [2] proposes an active shape model to describe the mouth area. The approach works first by training the deformations using spatiotemporal sets. Parameters of a Kalman Filter are defined at specific points and then used to integrate the information coming from the training. The mouth area is not automatically located, and the points are given for the training. Experiments are shown for tracking the mouth in some image sequences. Faruquie et. al. report an integrated audio-visual speech recognition system in [5]. They use active shape models with hand marked 46 knot points in the mouth area, and train the Point Distribution Models (PDM) in order to get the principal deformation modes. The lip contour is represented using five parabolic curves approximated by a fitting using the sets of 46 points. Features are computed for interior and exterior lips contours and then passed to classification. Highest figures achieved for video only recognition were reported as 31% success rate. A geometric lip model based on a quadratic curve is presented by Liew et. al. in [7] to use it as a complete model for lip detection from gray level images. An algorithm employing an stochastic cost function to find lip and non-lip regions in the images is then implemented in order to fit the model to images. Authors report detecting

the lip regions, although no recognition is mentioned after the regions is detected. Matthews [8] proposed variations on the active shape model to work into an integrated system of audio-visual speech recognition. A Multiscale Spatial Analysis (MSA) approach gave the best results achieving rates of 77% reported on a particular database of words. The approach is heavily dependent on the training of the parameters in order to find the correct locations in the scale histograms. Results of the work with MSA are also reported in [9]. Sadeghi et. al. in [11] proposed a model using Gaussian mixtures and general covariance matrix functions in order to estimate a mouth area from a set of examples. Results are given showing the approach is robust for segmenting mouth areas and lips for a set of gray level images. Neither tracking nor recognition is performed or mentioned in the work. As it can be seen from the literature on Visual Speech Recognition the proposed methods work mainly tracking the mouth of an specific speaker, and then extracting features for comparison with trained sequences of a finite vocabulary. In this paper we propose a method that first find the mouth and lips area in the image, estimate and extract special features for use in the classification stage, and in the experiments none of the mentioned stages are restricted to a specific speaker and pre-trained sequences. So, besides the original details of the method which will be given in the next sections, the conditions of the experimental setup and restrictions of our method are different from most of the work we found in the literature. This paper is further organized as follows. Section 2 describes the approach proposed here. Section 3 explains the experimentation setup for testing the approach and gives the results on classification. Section 4 outlines major conclusions as well as gives directions for future lines of work. 2. The Approach The approach we propose here is organized into three modules: 1) Segmentation of Lips Contours; 2) Lips Features and Extraction; 3) Recognition. Further sections explain in details these modules. 2.1 Segmentation of Lips Contours In order to provide an automatic solution for visual speech recognition the mouth area in an image sequence has to be detected, and the contouring of the lips segmented and followed apart from the background and the rest of the face. This is a difficult segmentation problem, particularly because of lighting conditions, speed of speech, and different aspects of someone s mouth such as teeth and lips area. A solution to work for more than one speaker will have to take those issues into consideration. We have designed a solution which is not aimed to extract the exact lips contours at each frame of the speech, but to provide stable points of the lips contours for the next stage of feature extraction. The input is a sequence of gray level images from the start until the end of a word pronouncing, and for segmentation of lips contours we propose the following steps: 1) Movement detection; 2) Enhancement filtering; 3) Entropy thresholding; 4) Mouth region detection; and 5) Lips contours identification. Movement detection works by computing a difference between two consecutive frames of a sequence, being a frame at time t minus an averaged filtered version of frame at time t-1. The output frame computed this way gives a more stable account region moved, Lie and Hsieh [6]. IEt 1 I t 1 * Mask (1) 1/ 9 1/ 9 1/ 9 where, Mask = 1/ 9 1/ 9 1/ 9 (2) 1/ 9 1/ 9 1/ 9 DI I t IEt 1 (3) (a) (b) (c) Figure 1..Consecutive frames of a sequence, and its compensated movement image. (a) Original frame at instant t-1, I t1 ; (b) Original frame at instant t, I t ; (c) Detected frame DI computed from (a) and (b). Figure 1 shows two consecutive frames and its computed new frame with movement compensated. It can be seen from Figure 1.(c) that lighting became more uniform and strong details were enhanced.

Enhancement filtering is designed to enhance the lips and mouth regions, and to diminish noise smaller than a structuring element. It works by applying to each detected frame DI a gray level morphological opening with the structuring element as in (4) with 8 iterations. Figure 2 shows a result of applying the process of enhancement. (a) (b) Figure 2 Result of morphological filtering to enhance lips and mouth region. (a) Original frame; (b) Enhanced image after application of (4). Entropy thresholding was a way we tested in order to remove the background of the image, and most of the face features in order to leave the mouth area as a more uniform region in the image. We have implemented the algorithm of Abutaleb [1], and applied it to each frame of the sequence. Figure 3 shows an application of such process in one frame. (4) (a) (b) Figure 4 Result for mouth detection. (a) Frame from last stage already in binary; (b) Result after applying mouth detection stage. Lips contour identification stage uses an structuring element to erode the mouth region, and then the eroded version minus the mouth image provides the lips contours. Cleaning for remaining artifacts found is then performed for final output of only a major lips contour, i.e. external. Figure 5 shows results for lips contour identification in one frame of our data. (a) (b) (c) Figure 5 Result of contour identification after mouth detection. (a) Mouth region from previous stage; (b) Result with lips contour. The image (b) is the negative since it is the direct result of the eroded minus the mouth image; (c) Final contour after chain coding is performed. (a) (b) Figure 3 Result after entropy thresholding. (a) Enhanced frame from last stage; (b) Image resultant after applying Abutaleb thresholding. Mouth region detection separates upon the uniform regions, already in binary from the last stage, the mouth from the rest of the image. The hypothesis that the mouth would be the largest region left works well at this stage. Figure 4 gives an example of mouth detection. 2.2 Lips Features and Extraction The output from the last stage of lips contours identification is passed to a chain code marking algorithm, Davies [4], in order to provide a more efficient data structure for feature extraction (see Figure 5 (c)). We have used eight directions in the code, which also cleared some jagging on the lips contours. Most solutions in the visual speech recognition, see Matthews [8] and Matthews et.al. [9], for recent examples, choose to approximate the lips contours using splines, or active contours, and perform tracking frame by frame of the lips movements. This solution has a high computational cost, and it usually needs pre-defined starting control points. We propose here in this work to use a different and much smaller set

of features for the lips than the closed 2-D contours. They are easier to extract, and more stable, and since the final aim is to perform the recognition, we tested them reaching comparable results with the top known published results nowadays. Since the significant differences that can be detected from someone s mouth while speaking a word are the opening of the mouth, and its speed, we propose as features for the lips four (4) distances: 1) H 1, the largest horizontal; 2) V 1, a vertical distance exactly at midpoint of H 1 ; 3) V 2, a vertical distance at midpoint of the left part of H 1 from the crossing of V 1 ; and 4) V 3, a vertical distance at midpoint of the right part of H 1 from the crossing of V 1. Figure 6 shows the four distances proposed as features schematically from the lips contours. 3. Experiments We have collected for the experiments 15 samples for each word of the vocabulary set. The words chosen are { zero, um, dois, três, quatro, cinco, seis, sete, oito, nove }, respectively related to the numerals {0,1,2,3,4,5,6,7,8,9}. The database includes more than twenty (20) different speakers pronouncing one, or more sequences. Figure 7 shows part of an original sequence of a speaker pronouncing the word um (1). The camera sampled at frame rate of 30 frames per second. Figure 8 shows the first three frames of the sequence on the top row and the last three on the bottom row. Figure 6 Image showing the features H 1, V 1, V 2, and V 3 to be extracted and used for visual speech recognition. This set of lips features are then computed for each frame of a sequence, and a feature vector to be passed for recognition is then a list of the number of frames times four features for each frame. 2.3 Words Classification based on the Visual Features The number of frames for the sequences of speakers is not constant, since even for the same word a different speaker would take longer to pronounce it. We did not want to constrain our solution to a fixed number of frames. For the words classification stage we compared all the feature vectors for different words and speakers, and the ones with less frames than others are filled with zeros in those particular frames missing. Nearest neighbor with Euclidean distance was used in order to classify the feature vector onto each of the ten classes proposed. Next section explains the experimental setup used and gives the main results achieved by the proposed method. Figure 7 Some input frames of a sequence pronouncing word 1 ( um ). Figure 8 shows respectively, for the same sequence from Figure 7, the results after the stages of movement detection, enhancement filtering and entropy thresholding. The images on Figure 8 are already in binary showing mainly the mouth region, nostrils, a part of background and still some noise. Figure 8 Frames of a sequence pronouncing word 1 ( um )., results after motion detection and morphological enhancement compensation. Figure 9 gives the outputs, for the sequence in Figure 8, after mouth region detection and lips contour

identification. It can be seen on the image frames in Figure 9 that there are more contours than the external lips, and this is the reason why it is needed cleaning for not closed contours left. Next stage, see the output at Figure 10, performs this cleaning and chain code the lips contours for further processing. Figure 11 Some input frames of a sequence pronouncing word 8 ( oito ). Figure 9 Frames of a sequence pronouncing word 1 ( um ), results after lips contours detection. Figure 12 Frames of a sequence pronouncing word 8 ( oito )., results after motion detection and morphological enhancement compensation. Figure 10 Frames of a sequence pronouncing word 1 ( um ), results with final contour detected. Figures 11, 12, 13 and 14 shows respectively frames of a sequence of word oito, being the original data, output at point of entropy thresholding, lips identification, and lips contours. We have chosen to show images of these two words in order for the reader to see the difficulty in classifying the speech based on the visual data at each frame. Our chosen features helped since they were more stable than using the whole contour. Figure 13 Frames of a sequence pronouncing word 8 ( oito ), results after lips contours detection.

0 1 2 3 4 5 6 7 8 9 0 0,05 0,00 0,00 0,22 0,20 0,15 0,13 0,00 0,02 0,05 1 0,00 0,97 0,03 0,00 0,00 0,00 0,00 0,00 0,00 0,00 2 0,00 0,00 0,93 0,00 0,00 0,00 0,07 0,00 0,00 0,00 3 0,10 0,00 0,00 0,07 0,13 0,00 0,38 0,00 0,32 0,00 4 0,02 0,00 0,00 0,22 0,25 0,08 0,18 0,03 0,15 0,07 5 0,02 0,00 0,00 0,00 0,10 0,42 0,15 0,22 0,0 0,10 6 0,03 0,00 0,05 0,38 0,13 0,05 0,17 0,02 0,17 0,00 Figure 14 Frames of a sequence pronouncing word 8 ( oito ), results with final contour detected. From a total of 150 samples, or sequences, being 15 for each word the system was run directly on the original data to extract the 150 feature vectors for classification. A typical sequence would have 45 frames, although this number varies depending on the word and the speaker for each sample. For testing the accuracy for the recognition of the words we chose randomly 20% of the samples and labeled them in order to computer the centers of the clusters for each word. The 80% left was then classified using nearest neighbor with Euclidean distance. A rank was made from the closest to the most distant. Five (5) batteries of tests were made following this procedure, always changing the samples from the 20%-80% subsets. This way we tested with all the samples into the two situations and averaged the results. Table 1 gives a confusion matrix for the ten (10) classes. It can be seen from the results on Table 1 that words um, dois, quatro, and cinco were classified correctly as first choices, and words seis and nove were classified as second choices. Those figures are encouraging since the conditions of testing were challenging because of different speakers used, automatic location of mouth and lips region, unfixed number of frames for the speech, and relative small sample size with difficult vocabulary (i.e. the words are either one, or two pitches apart, in the sense that you could separate the sound wave onto two major sounds). The recent literature in the area [3, 9] indicates that for an small size vocabulary, mostly tested using controlled conditions, and one speaker only, positive classifications figures of 40% were the highest. Considering the top four words classified correctly as shown in Table 1 the average of success is 64%, and for the complete set of experiments the success rate is 35%. 7 0,17 0,00 0,00 0,00 0,10 0,22 0,00 0,20 0,05 0,27 8 0,15 0,00 0,00 0,00 0,15 0,18 0,05 0,03 0,15 0,28 9 0,12 0,00 0,00 0,00 0,02 0,48 0,02 0,07 0,00 0,30 Table 1 Confusion matrix for the ten classes of words tested. Values indicate percentage of positive classification, and they were averaged using five batteries of tests in a hold-out scheme. 4. Conclusions and Future Work In this paper we have shown a novel solution we developed for visual speech recognition. The method embraces the tasks of automatic lips detection, contour and feature extraction, and recognition. More specifically the automatic lips contour detection and the features designed and tested for the set of words recognition are original contributions of this work. We have collected 150 samples using a digital camera with more than 20 speakers pronouncing the set of numerals from zero to nove in Brazilian Portuguese. The experimental results reached a top success rate of over 64% for the correctly classified words, and on average the success rate was 35%. Recent published results from the literature, Matthews in [8] indicate a range from 30-40% success rate at the state of the art systems. The results we achieved are promising, since there is room for improvement considering adding more features for the lips and maybe synchronizing the speech acquired. Visual speech recognition is an important area of application for dictation and interface systems, especially considering the new standards for digital video and television being adopted in the market place. Important to say that visual speech recognition is not meant to achieve recognition by its own, it is a rather complementary tool in the broader context of Audio-visual speech recognition. Besides being a novel and competitive approach the work presented here appears to be one of the first regarding Brazilian Portuguese evaluation. Lines of future work includes evaluating the approach with more data and

larger vocabulary, as well as designing visual features more sensitive to a specific set of words and visemes. We plan to use video data, as in TV broadcast news, as well as a vocabulary of 3K words in Brazilian Portuguese. For this extension we think that an HMM (Hidden Markov Model) would improve the recognition rate since it will be more precise to distinguish the movement and speed constraints, please see Rabiner [10] for more details on HMMs. Acknowledgements We thank the people who volunteered as anonymous speakers for the database we collected, and all friends from the former Laboratory for Intelligent Systems (LIS) (1998-2000). References [1] A. Abutaleb, Automatic Thresholding of Gray-level Pictures using Two-dimensional Entropy, Computer Graphics and Image Understanding,, 41(1) (1989), 22 32. [2] A. Caplier, Lip Detection and Tracking, in Proceedings of the 11 th Conference.. on Image Analysis and Processing (CIAP) (2001), IEEE CS Press, 8--13 [3] T. Chen, Audiovisual Speech Processing: lip reading and lip synchronization, IEEE Signal Processing Magazine, January (2001), 9 21. [4] E. Davies, Machine Vision: theory, algorithms, practicalities, 2 nd. Ed., Academic Press, USA, 1997. [5] T. Faruquie, A. Majundar, N. Rajput, and L. Subramaniam, Large Vocabulary Audio-Visual Speech Recognition using Active Shape Models, in Proceedings of the 15 th International Conference. on Pattern Recognition (ICPR) (2000), IEEE CS Press, 106 109. [6] W. Lie and H. Hsieh, Lip Detection by Morphological Image Processing, in Proceedings of International Conference. on Signal Processing (ICSP) (1998), 7--13. [7] A. Liew, S. Leung and W. Lau, Region-based Approach to Robust Lip Contour Extraction, Electronics Letters 22 (15)(2000), 1272--1274. [8] I. Matthews, Features for Audio-Visual Speech Recognition, Ph.D. Thesis, School of Information Systems, University of East Anglia, UK, 1998. [9] I. Matthews, T. Cootes, J. Bangham, S. Cox and R. Harvey, Extraction of Visual Features for Lipreading, IEEE Transactions on Pattern Analysis and Machine Intelligence 24 (2) (2002), 779--789. [10] L. R. Rabiner, A Tutorial on hidden Markov models and selected applications in speech recognition, Proceedings of the IEEE 77(2) (1989), 257 286. [11] M. Sadeghi, J. Kittler, and K. Messer, Modelling and Segmentation of Lip Area in Face Images, IEE Proceedings on Vision. Image, and Signal Processing 149(3) (2002), 179--184. [12] Q. Summerfield. Lipreading and audio-visual speech perception, Philosophical Transactions of the Royal Society of London B, (335), 71 78, 1992.