Pump-priming PASCAL proposal: Large Margin Algorithms and Kernel Methods for Speech Applications

Research on large margin algorithms in conjunctions with kernels methods has been both exciting and successful. While there have been quite a few preliminary successes in applying kernel methods for speech applications, most the research efforts have focused on non-temporal problems such as text classification and optical character recognition (OCR). We propose to design, analyze, and implement learning algorithms and kernels for hierarchical-temporal speech utterances. Our first and primary end-goal is to build and test thoroughly a full-blown speech phoneme classifier that will be trained on millions of examples and will achieve the best results in this domain. We also plan to apply and test the resulting algorithms and kernels to other supervised problems in spoken language such as language identification, word spotting from phonemes, and speaker verification. 1 Background and Motivation We propose an algorithmic research framework for supervised speech analysis problems that builds on recent advances in kernel methods and large margin classifiers. Specific applications include (but are not limited to) speech phoneme classification, spoken language identification, closed vocabulary word recognition/spotting, and speaker verification. Recent work on large margin methods such as support vector machines and boosting algorithms has shown to be effective in many decision tasks. While there has been some work on large margin methods for complex decision problems, most of the research focus still revolves around the less complex structures. Speech signals however exhibit a complex temporal structure in both the input space (acoustic signal) and the target space (e.g. phonetic transcription and word transcription). For instance, speech signals exhibit multi-scale temporal behavior and the set of speech phonemes is organized in an hierarchical structure. The current state-of-the-art models to handle such speech signals are based on generative models that capture some temporal dependencies such as Hidden Markov Models (HMMs). (See for instance [9] and the many references therein.) Numerous reasons have underscored HMMs as the tool of choice in speech signal processing, such as their ability to constrain the space of possible solutions to legal and probable sequences of phonemes/words, through the design of speech-specific Markovian state topologies, as well as their insensitivity to large unbalanced datasets due to their generative nature.