A multimodal pattern recognition framework for speaker detection

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.

Error handling in multimodal voice-enabled interfaces of tour-guide robots using graphical models

Plamen Prodanov

Mobile service robots are going to play an increasing role in the society of humans. Voice-enabled interaction with service robots becomes very important, if such robots are to be deployed in real-world environments and accepted by the vast majority of potential human users. The research presented in this thesis addresses the problem of speech recognition integration in an interactive voice-enabled interface of a service robot, in particular a tour-guide robot. The task of a tour-guide robot is to engage visitors to mass exhibitions (users) in dialogue providing the services it is designed for (e.g. exhibit presentations) within a limited time. In managing tour-guide dialogues, extracting the user goal (intention) for requesting a particular service at each dialogue state is the key issue. In mass exhibition conditions speech recognition errors are inevitable because of noisy speech and uncooperative users of robots with no prior experience in robotics. They can jeopardize the user goal identification. Wrongly identified user goals can lead to communication failures. Therefore, to reduce the risk of such failures, methods for detecting and compensating for communication failures in human-robot dialogue are needed. During the short-term interaction with visitors, the interpretation of the user goal at each dialogue state can be improved by combining speech recognition in the speech modality with information from other available robot modalities. The methods presented in this thesis exploit probabilistic models for fusing information from speech and auxiliary modalities of the robot for user goal identification and communication failure detection. To compensate for the detected communication failures we investigate multimodal methods for recovery from communication failures. To model the process of modality fusion, taking into account the uncertainties in the information extracted from each input modality during human-robot interaction, we use the probabilistic framework of Bayesian networks. Bayesian networks are graphical models that represent a joint probability function over a set of random variables. They are used to model the dependencies among variables associated with the user goals, modality related events (e.g. the event of user presence that is inferred from the laser scanner modality of the robot), and observed modality features providing evidence in favor of these modality events. Bayesian networks are used to calculate posterior probabilities over the possible user goals at each dialogue state. These probabilities serve as a base in deciding if the user goal is valid, i.e. if it can be mapped into a tour-guide service (e.g. exhibit presentation) or is undefined – signaling a possible communication failure. The Bayesian network can be also used to elicit probabilities over the modality events revealing information about the possible cause for a communication failure. Introducing new user goal aspects (e.g. new modality events and related features) that provide auxiliary information for detecting communication failures makes the design process cumbersome, calling for a systematic approach in the Bayesian network modelling. Generally, introducing new variables for user goal identification in the Bayesian networks can lead to complex and computationally expensive models. In order to make the design process more systematic and modular, we adapt principles from the theory of grounding in human communication. When people communicate, they resolve understanding problems in a collaborative joint effort of providing evidence of common shared knowledge (grounding). We use Bayesian network topologies, tailored to limited computational resources, to model a state-based grounding model fusing information from three different input modalities (laser, video and speech) to infer possible grounding states. These grounding states are associated with modality events showing if the user is present in range for communication, if the user is attending to the interaction, whether the speech modality is reliable, and if the user goal is valid. The state-based grounding model is used to compute probabilities that intermediary grounding states have been reached. This serves as a base for detecting if the the user has reached the final grounding state, or wether a repair dialogue sequence is needed. In the case of a repair dialogue sequence, the tour-guide robot can exploit the multiple available modalities along with speech. For example, if the user has failed to reach the grounding state related to her/his presence in range for communication, the robot can use its move modality to search and attract the attention of the visitors. In the case when speech recognition is detected to be unreliable, the robot can offer the alternative use of the buttons modality in the repair sequence. Given the probability of each grounding state, and the dialogue sequence that can be executed in the next dialogue state, a tour-guide robot has different preferences on the possible dialogue continuation. If the possible dialogue sequences at each dialogue state are defined as actions, the introduced principle of maximum expected utility (MEU) provides an explicit way of action selection, based on the action utility, given the evidence about the user goal at each dialogue state. Decision networks, constructed as graphical models based on Bayesian networks are proposed to perform MEU-based decisions, incorporating the utility of the actions to be chosen at each dialogue state by the tour-guide robot. These action utilities are defined taking into account the tour-guide task requirements. The proposed graphical models for user goal identification and dialogue error handling in human-robot dialogue are evaluated in experiments with multimodal data. These data were collected during the operation of the tour-guide robot RoboX at the Autonomous System Lab of EPFL and at the Swiss National Exhibition in 2002 (Expo.02). The evaluation experiments use component and system level metrics for technical (objective) and user-based (subjective) evaluation. On the component level, the technical evaluation is done by calculating accuracies, as objective measures of the performance of the grounding model, and the resulting performance of the user goal identification in dialogue. The benefit of the proposed error handling framework is demonstrated comparing the accuracy of a baseline interactive system, employing only speech recognition for user goal identification, and a system equipped with multimodal grounding models for error handling.

EPFL2006

Using auxiliary sources of knowledge for automatic speech recognition

Mathew Magimai Doss

Standard hidden Markov model (HMM) based automatic speech recognition (ASR) systems usually use cepstral features as acoustic observation and phonemes as subword units. Speech signal exhibits wide range of variability such as, due to environmental variation, speaker variation. This leads to different kinds of mismatch, such as, mismatch between acoustic features and acoustic models or mismatch between acoustic features and pronunciation models (given the acoustic models). The main focus of this work is on integrating auxiliary knowledge sources into standard ASR systems so as to make the acoustic models more robust to the variabilities in the speech signal. We refer to the sources of knowledge that are able to provide additional information about the sources of variability as auxiliary sources of knowledge. The auxiliary knowledge sources that have been primarily investigated in the present work are auxiliary features and auxiliary subword units. Auxiliary features are secondary source of information that are outside of the standard cepstral features. They can be estimation from the speech signal (e.g., pitch frequency, short-term energy and rate-of-speech), or additional measurements (e.g., articulator positions or visual information). They are correlated to the standard acoustic features, and thus can aid in estimating better acoustic models, which would be more robust to variabilities present in the speech signal. The auxiliary features that have been investigated are pitch frequency, short-term energy and rate-of-speech. These features can be modelled in standard ASR either by concatenating them to the standard acoustic feature vectors or by using them to condition the emission distribution (as done in gender-based acoustic modelling). We have studied these two approaches within the framework of hybrid HMM/artificial neural networks based ASR, dynamic Bayesian network based ASR and TANDEM system on different ASR tasks. Our studies show that by modelling auxiliary features along with standard acoustic features the performance of the ASR system can be improved in both clean and noisy conditions. We have also proposed an approach to evaluate the adequacy of the baseform pronunciation model of words. This approach allows us to compare between different acoustic models as well as to extract pronunciation variants. Through the proposed approach to evaluate baseform pronunciation model, we show that the matching and discriminative properties of single baseform pronunciation can be improved by integrating auxiliary knowledge sources in standard ASR. Standard ASR systems use usually phonemes as the subword units in a Markov chain to model words. In the present thesis, we also study a system where word models are described by two parallel chains of subword units: one for phonemes and the other are for graphemes (phoneme-grapheme based ASR). Models for both types of subword units are jointly learned using maximum likelihood training. During recognition, decoding is performed using either or both of the subword unit chains. In doing so, we thus have used graphemes as auxiliary subword units. The main advantage of using graphemes is that the word models can be defined easily using the orthographic transcription, thus being relatively noise free as compared to word models based upon phoneme units. At the same time, there are drawbacks to using graphemes as subword units, since there is a weak correspondence between the grapheme and the phoneme in languages such as English. Experimental studies conducted for American English on different ASR tasks have shown that the proposed phoneme-grapheme based ASR system can perform better than the standard ASR system that uses only phonemes as its subword units. Furthermore, while modelling context-dependent graphemes (similar to context-dependent phonemes), we observed that context-dependent graphemes behave like phonemes. ASR studies conducted on different tasks showed that by modelling context-dependent graphemes only (without any phonetic information) performance competitive to the state-of-the-art context-dependent phoneme-based ASR system can be obtained.

EPFL2005

AM-FM DECOMPOSITION OF SPEECH SIGNAL: APPLICATIONS FOR SPEECH PRIVACY AND DIAGNOSIS

Sriram Ganapathy, Hynek Hermansky, Petr Motlicek, Amrutha Prasad

Although current trends in speech processing consider deep learning through data-driven technologies, many potential applications exhibit lack of training or development data. Therefore, considerably light signal processing techniques are still of interest. This paper describes an efficient technique for decomposing the AM and FM components of the speech signal, which is not based on frame-by-frame short-time analysis of the signal. Instead, we estimate all-pole models of frequency-localized Hilbert envelopes of large segments of speech signal at different frequencies. The technique on decomposition of speech signal into AM and FM components appears to be of interest in voice studies benefiting from alleviation of the message-bearing components of speech (e.g. security oriented applications such as speaker recognition, or speech diagnosis often relying on spectra averaging to discard the content of the speech). Similarly, discarding speaker information while preserving the message in the speech is of interest for privacy-oriented applications. experimental results on automatic speech and speaker recognition tasks clearly show that the AM component preserves the content (message) of the speech, while the FM component carries the information related to the speaker.