Discovery of syllabic percussion patterns in tabla solo recordings

We address the unexplored problem of percussion pattern discovery in Indian art music. Percussion in Indian art music uses onomatopoeic oral mnemonic syllables for the transmission of repertoire and technique. This is utilized for the task of percussion pattern discovery from audio recordings. From a parallel corpus of audio and expert curated scores for 38 tabla solo recordings, we use the scores to build a set of most frequent syllabic patterns of different lengths. From this set, we manually select a subset of musically representative query patterns. To discover these query patterns in an audio recording, we use syllable-level hidden Markov models (HMM) to automatically transcribe the recording into a syllable sequence, in which we search for the query pattern instances using a Rough Longest Common Subsequence (RLCS) approach. We show that the use of RLCS makes the approach robust to errors in automatic transcription, significantly improving the pattern recall rate and F-measure. We further propose possible enhancements to improve the results.