Speech and text retrieval using phoneme sequence representation and continuous dynamic programming

摘要

A retrieval system between speech and text or speech and speech is proposed. Speech wave is transformed into a linear sequence of vector phoneme symbol obtained from frame-wise recognition of phoneme category. Each vector represents n-th candidates of phoneme recognition at t-frame. A text data is also transformed into a hear sequence of phoneme symbol with duplicated same phoneme symbols using its average duration. Spotting retrieval is realized by applying Continuous Dynamic Programming to these two kinds of phoneme symbol sequence. Some experiment results of speech retrieval from speech databases are shown to confirm the usefulness of the proposed method.