Speech Retrieval using Phonemes with Error Correction

Abstract

In speech retrieval, text queries are matched to spoken-word recordings. We undertook a series of retrieval experiments on such speech documents, using both phonemes and words for matching text to speech. We found that, given a reliable automatic word-based process for speech recognition, phoneme-based retrieval can be as effective as word-based retrieval. However, retrieval is v.!orse with phoneme-based speech recognition. Error correction techniques, such as manual error correction of phoneme sequences and use of a standard string edit distance, did not improve the effectiveness of phoneme-based retrieval.