Phoneme-Lattice to Phoneme-Sequence Matching Algorithm Based on Dynamic Programming

A novel phoneme-lattice to phoneme-sequence matching algorithm based on dynamic programming is presented in this paper. Phoneme lattices have been shown to be a good choice to encode in a compact way alternative decoding hypotheses from a speech recognition system. These are typically used for the spoken term detection and keyword-spotting tasks, where a phoneme sequence query is matched to a reference lattice. Most current approaches suffer from a lack of flexibility whenever a match allowing phoneme insertions, deletions and substitutions is to be found. We introduce a matching approach based on dynamic programming, originally proposed for Minimum Bayes decoding on speech recognition systems. The original algorithm is extended in several ways. First, a self-trained phoneme confusion matrix for phoneme comparison is applied as phoneme penalties. Also, posterior probabilities are computed per arc, instead of likelihoods and an acoustic matching distance is combined with the edit distance at every arc. Finally, total matching scores are normalized based on the length of the optimum alignment path. The resulting algorithm is compared to a state-of-the-art phoneme-lattice-to-string matching algorithm showing relative precision improvements over 20% relative on an isolated word retrieval task.