Speech fragment decoding techniques using silent pause detection

Abstract

Silent pause frequently occurs in spontaneous speech. When recognizing spontaneous speech, silent pause tends to degrade the performance of typical speech recognizers. This paper proposes a fragment decoding method to improve the performance of speech recognizer using silent pause detection. This method automatically detects silent pauses and cuts long utterance into speech fragments. At decoding stage, instead of being skipped, these silent fragments are decoded separately. Final transcription of the whole utterance can be derived from corresponding fragmental results. Further improvement is made to reduce the run-time consumed on decoding. Because of an introduction of accurate word boundary, the misrecognition at silent frames is declined. Recognition experiments conducted on monolog speech in tourism field show that the proposed method outperforms the traditional frame skipping method. © 2012 Springer-Verlag.