K Eyword S Potting on W Ord L Attices

In spite of its numerous potential applications, Automatic Speech Recognition (ASR) remains a difficult (and mainly unsolved) problem. In addition to the intrinsic difficulty of the task, users tend to go beyond the pre-defined lexicon words, and the important keywords necessary to understand voice requests are often lost in extra words. In this context, it is often interesting to develop Keyword Spotting (KWS) approaches that will focus on the detection and recognition of pre-defined keywords lost in unconstrained, conversational, speech. The goal of this project is to perform confidence based keyword spotting on word lattices, which are a compact way for storing the most probable hypotheses previously generated by a Large Vocabulary Continuous Speech Recognizer (LVCSR). Every spoken utterance is turned into a word lattice, and fast search and rescoring techniques are then applied to detect keywords in the lattice, using their confidence scores to take an accept/reject decision. By doing so, more knowledge (lexical and syntactic) can be taken into account in the first pass (LVCSR pass), while limiting the search space in the second pass, hence also allowing us to use more complex algorithms. More specific to the work presented here, keyword hypotheses posteriors are estimated from the reduced search space (the lattices), and used as confidence scores to take the final decision. As posteriors minimize, by definition, the probability of error, we thus aim at minimizing the false alarm rate while maximizing true detection rates. Moreover, in order to take into account the word posterior probability mass dispersion among parallel lattice edges, various posterior rescoring techniques are investigated, all based on posterior accumulation over keyword hypotheses. Generation and rescoring of the above word lattices were all based on the popular HiddenMarkov Model (HMM) approaches. In this context, two particular instances of HMMs have been used and compared; the standard approach using Gaussian Mixture Models (GMMs) to estimate local (likelihood) scores, and a second one using a Multilayer Perceptron (MLP) directly estimating local posteriors.