This paper proposes a new method for spotting phrases in continuous speech. The phrase spotting is done based on the weighted sum of an acoustic likelihood and a prosodic phrase boundary likelihood. The prosodic phrase boundary likelihood is calculated statistically using pitch pattern HMMs. Introducing the prosodic phrase boundary likelihood is considered to have an effect to suppress the false alarms, because if the start and end points of the phrase candidate are different from the correct phrase boundaries the spotting score is lowered. The method was evaluated by Japanese phrase spotting experiments. The results show that the phrase boundary likelihood is calculated reasonably, and the phrase detection rate is improved using the prosodic phrase boundary likelihood.
Bibliographic reference. Hanazawa, Toshiyuki / Abe, Yoshiharu / Nakajima, Kunio (1995): "Phrase spotting using pitch pattern information", In EUROSPEECH-1995, 2137-2140.