Visual speech recognition | EPFL Graph Search

Speech is the most natural means of communication for humans. Therefore, since the beginning of computers it has been a goal to interact with machines via speech. While there have been gradual improvements in this field over the decades, and with recent drastic progress more and more commercial software is available that allow voice commands, there are still many ways in which it can be improved.

One way to do this is with visual speech information, more specifically, the visible articulations of the mouth. Based on the information contained in these articulations, visual speech recognition (VSR) transcribes an utterance from a video sequence. It thus helps extend speech recognition from audio-only to other scenarios such as silent or whispered speech (e.g.\ in cybersecurity), mouthings in sign language, as an additional modality in noisy audio scenarios for audio-visual automatic speech recognition, to better understand speech production and disorders, or by itself for human machine interaction and as a transcription method.

In this thesis, we present and compare different ways to build systems for VSR: We start with the traditional hidden Markov models that have been used in the field for decades, especially in combination with handcrafted features. These are compared to models taking into account recent developments in the fields of computer vision and speech recognition through deep learning. While their superior performance is confirmed, certain limitations with respect to computing power for these systems are also discussed.

This thesis also addresses multi-view processing and fusion, which is an important topic for many current applications. This is due to the fact that a single camera view often cannot provide enough flexibility with speakers moving in front of the camera. Technology companies are willing to integrate more cameras into their products, such as cars and mobile devices, due to lower hardware cost for both cameras and processing units, as well as the availability of higher processing power and high performance algorithms. Multi-camera and multi-view solutions are thus becoming more common, which means that algorithms can benefit from taking these into account. In this work we propose several methods of fusing the views of multiple cameras to improve the overall results.

We can show that both, relying on deep learning-based approaches for feature extraction and sequence modelling, as well as taking into account the complementary information contained in several views, improves performance considerably. To further improve the results, it would be necessary to move from data recorded in a lab environment, to multi-view data in realistic scenarios. Furthermore, the findings and models could be transferred to other domains such as audio-visual speech recognition or the study of speech production and disorders.

Graph-based image representation learning

Renata Khasanova

Though deep learning (DL) algorithms are very powerful for image processing tasks, they generally require a lot of data to reach their full potential. Furthermore, there is no straightforward way to impose various properties, given by the prior knowledge about the target task, on the features extracted by a DL model. Therefore, in this thesis we propose several techniques that rely on the power of graph representations to embed prior knowledge inside the learning process. This allows to reduce the solution space and leads to faster optimization convergence and higher accuracy in the representation learning. In our first work, inspired by the ability of a human to correctly classify rotated, shifted or flipped objects, we propose an algorithm that permits to inherently encode invariance to isometric transformations of objects in an image. Our DL architecture is based on graph representations and consists of three novel layers, which we refer to as graph convolutional, dynamic pooling and statistical layers. Our experiments on the image classification tasks show that our network correctly recognizes isometrically transformed objects even though such types of transformation are not seen by the network at training time. Standard DL techniques are typically not able to succeed in solving such a problem without extensive data augmentation. Then, we propose to exploit the properties of graph-based approaches to efficiently process images with various types of projective geometry. In particular, we are interested in increasingly popular omnidirectional cameras, which have a 360 degree field of view. Despite their effectiveness, such cameras create images with specific geometric properties, which require special techniques for efficient processing. We propose an efficient way of adjusting the weights of the graph edges to adapt the filter responses to the geometric image properties introduced by omnidirectional cameras. Our experiments prove that using the proposed graph with properly adjusted edge weights permits to reach better performance as compared to using regular grid graph with equal weights. Finally, the approach described above relies on the isotropic filters, which work well within our transformation invariant architecture for image classification. However, for other problems (e.g. image compression) or even when used without dynamic pooling and statistical layers that are defined within the proposed architecture, these filters are unable to efficiently encode the information about the object. Thus, we introduce a different technique based on anisotropic filters that adapt their shape and size according to the omnidirectional image geometry. The main advantage of this approach compared to the previous one is the ability to encode the orientation of an image pattern, which is important for various tasks such as image compression. Our experiments show that our approach adapts to different image projective geometries and achieves state-of-the-art performance on image classification and compression tasks. Overall we propose several methods, which combine the power of DL and graph signal processing towards incorporating prior information about the target task inside the optimization procedure. We hope that the research efforts presented in this thesis will help the development of efficient DL algorithms that can use various types of prior knowledge to make them efficient even when the available training data is scarce.

EPFL2019

Prior knowledge in Kernel methods

Alexei Pozdnoukhov

Machine Learning is a modern and actively developing field of computer science, devoted to extracting and estimating dependencies from empirical data. It combines such fields as statistics, optimization theory and artificial intelligence. In practical tasks, the general aim of Machine Learning is to construct algorithms able to generalize and predict in previously unseen situations based on some set of examples. Given some finite information, Machine Learning provides ways to exract knowledge, describe, explain and predict from data. Kernel Methods are one of the most successful branches of Machine Learning. They allow applying linear algorithms with well-founded properties such as generalization ability, to non-linear real-life problems. Support Vector Machine is a well-known example of a kernel method, which has found a wide range of applications in data analysis nowadays. In many practical applications, some additional prior knowledge is often available. This can be the knowledge about the data domain, invariant transformations, inner geometrical structures in data, some properties of the underlying process, etc. If used smartly, this information can provide significant improvement to any data processing algorithm. Thus, it is important to develop methods for incorporating prior knowledge into data-dependent models. The main objective of this thesis is to investigate approaches towards learning with kernel methods using prior knowledge. Invariant learning with kernel methods is considered in more details. In the first part of the thesis, kernels are developed which incorporate prior knowledge on invariant transformations. They apply when the desired transformation produce an object around every example, assuming that all points in the given object share the same class. Different types of objects, including hard geometrical objects and distributions are considered. These kernels were then applied for images classification with Support Vector Machines. Next, algorithms which specifically include prior knowledge are considered. An algorithm which linearly classifies distributions by their domain was developed. It is constructed such that it allows to apply kernels to solve non-linear tasks. Thus, it combines the discriminative power of support vector machines and the well-developed framework of generative models. It can be applied to a number of real-life tasks which include data represented as distributions. In the last part of the thesis, the use of unlabelled data as a source of prior knowledge is considered. The technique of modelling the unlabelled data with a graph is taken as a baseline from semi-supervised manifold learning. For classification problems, we use this apporach for building graph models of invariant manifolds. For regression problems, we use unlabelled data to take into account the inner geometry of the input space. To conclude, in this thesis we developed a number of approaches for incorporating some prior knowledge into kernel methods. We proposed invariant kernels for existing algorithms, developed new algorithms and adapted a technique taken from semi-supervised learning for invariant learning. In all these cases, links with related state-of-the-art approaches were investigated. Several illustrative experiments were carried out on real data on optical character recognition, face image classification, brain-computer interfaces, and a number of benchmark and synthetic datasets.

EPFL2006

A multimodal pattern recognition framework for speaker detection

Patricia Besson

Speaker detection is an important component of a speech-based user interface. Audiovisual speaker detection, speech and speaker recognition or speech synthesis for example find multiple applications in human-computer interaction, multimedia content indexing, biometrics, etc. Generally speaking, any interface which relies on speech for communication requires an estimate of the user's speaking state (i.e. whether or not he/she is speaking to the system) for its reliable functioning. One needs therefore to identify the speaker and discriminate from other users or background noise. A human observer would perform such a task very easily, although this decision results from a complex cognitive process referred to as decision-making. Generally speaking, this process starts with the acquisition by the human being of information about the environment, through each of its five senses. The brain then integrates these multiple information. An amazing property of this multi-sensory integration by the brain, as pointed out by cognitive sciences, is the perception of stimuli of different modalities as originating from a single source, provided they are synchronized in space and time. A speaker is a bimodal source emitting jointly an auditory signal and a visual signal (the motion of the articulators during speech production). The two signals are obviously co-occurring spatio-temporally. This interesting property allows us – as human observers – to discriminate between a speaking mouth and a mouth whose motion is not related with the auditory signal. This dissertation deals with the modelling of such a complex decision-making, using a pattern recognition procedure. A pattern recognition process comprises all the stages of an investigation, from data acquisition to classification and assessment of the results. In the audiovisual speaker detection problem, tackled more specifically in this thesis, the data are acquired using only one microphone and camera. The pattern recognizer integrates and combines these two modalities to perform and is therefore denoted as "multimodal". This multimodal approach is expected to increase the performance of the system. But it also raises many questions such as what should be fused, when in the decision process this fusion should take place, and how is it to be achieved. This thesis provides answers to each of these issues through the proposition of detailed solutions for each step of the classification process. The basic principle is to evaluate the synchrony between the audio and video features extracted from potentially speaking mouths, in order to classify each mouth as speaking or not. This synchrony is evaluated through a mutual information based function. A key to success is the extraction of suitable features. The audiovisual data are then processed through an information theoretic feature extraction framework after having been acquired and represented in a tractable way. This feature extraction framework uses jointly the two modalities in a feature-level fusion scheme. This way, the information originating from the common source is recovered while the independent noise is discarded. This approach is shown to minimize the probability of committing an error on the source estimate. These optimal features are put as inputs of the classifier, defined through a hypothesis testing approach. Using jointly the two modalities, it outputs a single decision about the class label of each candidate mouth region ("speaker" or "non-speaker"). Therefore, the acoustic and visual information are combined at both the feature and the decision levels, so that we can talk about a hybrid fusion method. The hypothesis testing approach gives means for evaluating the performance of the classifier itself but also of the whole pattern recognition system. In particular, the added-value offered by the feature extraction step can be assessed. The framework is applied in a first time with a particular emphasis on the audio modality: the information theoretic feature extraction addresses the optimization of the audio features using jointly the video information. As a result, audio features specific to speech production are produced. The system evaluation framework establishes that putting these features at input of the classifier increases its discrimination power with respect to equivalent non-optimized features. Then the enhancement of the video content is addressed more specifically. The mouth motion is obviously the suitable video representation for handling a task such as speaker detection. However, only an estimate of this motion, the optical flow, can be obtained. This estimation relies on the intensity gradient of the image sequence. Graph theory is used to establish a probabilistic model of the relationships between the audio, the motion and the image intensity gradient, in the particular case of a speaking mouth. The interpretation of this model leads back to the optimization function defined for the information theoretic feature extraction. As a result, a scale-space approach is proposed for estimating the optical flow, where the strength of the smoothness constraint is controlled via a mutual information based criterion involving both the audio and the video information. First results are promising even if more extensive tests should be carried out, in noisy conditions in particular. As a conclusion, this thesis proposes a complete pattern recognition framework dedicated to audiovisual speaker detection and minimizing the probability of misclassifying a mouth as "speaker" or "non-speaker". The importance of fusing the audio and video content as soon as at the feature level is demonstrated through the system evaluation stage included in the pattern recognition process.

EPFL2007