Search by lyrics, the problem of locating the exact occurrences of a phrase from lyrics in musical audio, is a recently emerging research topic. Unlike key-phrases in speech, lyrical key-phrases have durations that bear important relation to other musical aspects like the structure of a composition. In this work we propose an approach that address the differences of syllable durations, specific for singing. First a phrase is expanded to MFCC-based phoneme models, trained on speech. Then, we apply ...
Search by lyrics, the problem of locating the exact occurrences of a phrase from lyrics in musical audio, is a recently emerging research topic. Unlike key-phrases in speech, lyrical key-phrases have durations that bear important relation to other musical aspects like the structure of a composition. In this work we propose an approach that address the differences of syllable durations, specific for singing. First a phrase is expanded to MFCC-based phoneme models, trained on speech. Then, we apply dynamic time warping between the phrase and audio to estimate candidate audio segments in the given audio recording. Next, the retrieved audio segments are ranked by means of a novel score-informed hidden Markov model, in which durations of the syllables within a phrase are explicitly modeled. The proposed approach is evaluated on 12 a-capella audio recordings of Turkish Makam music. Relying on standard speech phonetic models, we arrive at promising results that outperform a baseline approach unaware of lyrics durations. To the best of our knowledge, this is the first work tackling the problem of search by lyrical key-phrases. We expect that it can serve as a baseline for further research on singing material with similar musical characteristics.
+
