Breadth-first search for finding the optimal phonetic transcription from multiple utterances.

摘要: Extending the vocabulary of a large speech recognition system usually requires phonetic transcriptions for all words to be known. With automatic baseform determination acoustic samples in question can substitute required expert knowledge. In this paper we follow probabilitistic approach problem and present novel breadth-first search algorithm which takes full advantage multiple samples. An extension genereate phone graphs as well an EM based iteration scheme estimating stochastic pronunciation models is presented. preliminary experiments phoneme error rates below 5% with respect standard are achieved without language or word specific prior