PHAST: Spoken Document Retrieval Based on Sequence Alignment

This paper presents a new approach to spoken document information retrieval for spontaneous speech corpora. Classical approach to this problem is the use of an automatic speech recognizer (ASR) combined with standard information retrieval techniques, based on terms or n-grams. However, state-of-the-art large vocabulary continuous ASRs produce transcripts of spontaneous speech with a word error rate of 25% or higher, which is a drawback for retrieval techniques based on terms or n-grams. In order to overcome such a limitation, our method is based on a sequence alignment algorithm drawn from the field of bioinformatics to search “sounds like” sequences in the document collection. These matching sequences are potentially misrecognized words from the ASR and can be used to retrieve relevant passages and documents from the collection. Our approach doesn’t depend on extra information provided by the ASR. We have evaluated and compared our approach to others in the state of the art in both spoken document retrieval and spoken passage retrieval tasks. The evaluation has been performed in the context of Question Answering using a corpus of automatic transcripts from the Spanish and European parliaments. The results show that our method outperforms by 10 points traditional term based search and n-gram search on automatic transcripts.