| Summary: | 碩士 === 國立清華大學 === 資訊工程學系 === 98 === This paper proposes an improved query by singing/humming (QBSH) system using both melody and lyrics information for achieving better performance. Singing/humming discrimination (SHD) is first performed to distinguish singing from humming queries. For a humming query, we apply a pitch-only melody recognition method that has been used for QBSH task at MIREX with rank-1 per-formance. For a singing query, we combine the scores from melody recognition and lyrics recognition to take advantage of the extra lyrics information. Lyrics recognition is based on a modified tree lexicon that is commonly used in speech recognition. The performance of the overall QBSH system achieves 39.01% and 23.53% error reduction rates, respectively, for top-20 recognition under two experimental settings, indicating the feasibility of the proposed method.
|
|---|
