Speech and Song Search on the Web: System Design and Implementation

This paper proposes a novel search system for speech and song segments. The amount of accumulated video data in the World Wide Web is expanding and its content is varied. Video content includes natural voices and singing voices, and these differ in their phoneme lengths. Our system uses frame-wise phoneme recognition and continuous dynamic programming (CDP). First, each target and query waveform is divided into fixed short-time frames; second, each frame of the waveform is used to estimate a phoneme label using Bayes estimation; third, the query sequences of phoneme labels are searched from target sequences by time-robustness CDP; and, finally, this system gets candidate answers. This method is robust along the time dimension, and thus has a great advantage for natural voice as well as song. This paper also introduces an implementation of this system, which is published on the Web, as a secondary search engine for Youtube data.