A passive approach to evaluating Mridangam transcription via perceptual experiment

Analysis and characterization of percussive instruments using computational methods have recently got attention both in the context of Western and non-Western repertoire, specifically on automatic transcription of tabla and mridangam strokes (Anantapadmanabhan et al., 2014; Guedes et al., 2018). The transcription approach, however, is limited by its dependency on prior knowledge about the specific modes of the instrument. Another concern is the absence of a unique mapping between the strokes and their nomenclature in the vocabulary. The same stroke is often uttered differently in the konakkol vocalization based on contextual variations or grammatical impositions. Most notably, even an expert musician is often unable to resolve such ambiguities on isolated presentation of a stroke but uses contextual cues to possibly interpolate within and across phrases. Our previous work (Ganguli et al., 2020; 2021) addressed this problem by proposing a combination of acoustic and semantic approaches for the contextual transcription of mridangam strokes. Grammatically-accurate labels for the transcribed strokes were obtained through the learned language model as knowledge-constraints.