Two-stage vocabulary-free spoken document retrieval - subword identification and re-recognition of the identified sections

A query word for retrieval systems is liable to be a special term not included in a speech recognizer dictionary. Spoken document retrieval (SDR) systems must therefore be vocabulary-free to deal with arbitrary query words. This paper proposes a new method for vocabulary-free spoken document retrieval. The method exploits two-stage tactics. First, when a query word is submitted, the query word is transformed to a subword sequence according to conversion rules. The subword sequence is searched for spoken documents previously transcribed to a subword sequence by subword recognition. The identified sections are extracted according to the distance between the subword sequences of the query and the identified sections. Second, each identified section is re-recognized using a grammar that includes the query subword sequence. Retrieval experiments were conducted with an actual TV program and the results demonstrated that the proposed method improved SDR performance without long delays in retrieval.