Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Weakly Supervised Tabla Stroke Transcription via TI-SDRM: A Rhythm-Aware Lattice Rescoring Framework

Rahul Bapusaheb Kodag, Vipul Arora

TL;DR

This work tackles weakly supervised Tabla Stroke Transcription by fusing a CTC-based acoustic model with a rhythm-aware TI-SDRM that simultaneously captures long-term tala structure and short-term adaptive dynamics. A latent tala posterior and a dynamic Dirichlet-multinomial rhythm model drive adaptive, history-sensitive lattice rescoring, implemented via state-expansion in an expanded decoding lattice. The authors curate real-world and synthetic tabla datasets and demonstrate consistent, substantial improvements in stroke error rate over acoustic-only decoding across multiple settings, including low-resource scenarios. The NN-LM provides a data-driven alternative to the static prior, with data efficiency favoring the Bayesian prior at small data scales. This framework advances weakly supervised transcription and offers a benchmark for rhythm-aware sequence correction in Hindustani percussion.

Abstract

Tabla Stroke Transcription (TST) is central to the analysis of rhythmic structure in Hindustani classical music, yet remains challenging due to complex rhythmic organization and the scarcity of strongly annotated data. Existing approaches largely rely on fully supervised learning with onset-level annotations, which are costly and impractical at scale. This work addresses TST in a weakly supervised setting, using only symbolic stroke sequences without temporal alignment. We propose a framework that combines a CTC-based acoustic model with sequence-level rhythmic rescoring. The acoustic model produces a decoding lattice, which is refined using a \textbf{$T\bar{a}la$}-Independent Static--Dynamic Rhythmic Model (TI-SDRM) that integrates long-term rhythmic structure with short-term adaptive dynamics through an adaptive interpolation mechanism. We curate a new real-world tabla solo dataset and a complementary synthetic dataset, establishing the first benchmark for weakly supervised TST in Hindustani classical music. Experiments demonstrate consistent and substantial reductions in stroke error rate over acoustic-only decoding, confirming the importance of explicit rhythmic structure for accurate transcription.

Weakly Supervised Tabla Stroke Transcription via TI-SDRM: A Rhythm-Aware Lattice Rescoring Framework

TL;DR

This work tackles weakly supervised Tabla Stroke Transcription by fusing a CTC-based acoustic model with a rhythm-aware TI-SDRM that simultaneously captures long-term tala structure and short-term adaptive dynamics. A latent tala posterior and a dynamic Dirichlet-multinomial rhythm model drive adaptive, history-sensitive lattice rescoring, implemented via state-expansion in an expanded decoding lattice. The authors curate real-world and synthetic tabla datasets and demonstrate consistent, substantial improvements in stroke error rate over acoustic-only decoding across multiple settings, including low-resource scenarios. The NN-LM provides a data-driven alternative to the static prior, with data efficiency favoring the Bayesian prior at small data scales. This framework advances weakly supervised transcription and offers a benchmark for rhythm-aware sequence correction in Hindustani percussion.

Abstract

Tabla Stroke Transcription (TST) is central to the analysis of rhythmic structure in Hindustani classical music, yet remains challenging due to complex rhythmic organization and the scarcity of strongly annotated data. Existing approaches largely rely on fully supervised learning with onset-level annotations, which are costly and impractical at scale. This work addresses TST in a weakly supervised setting, using only symbolic stroke sequences without temporal alignment. We propose a framework that combines a CTC-based acoustic model with sequence-level rhythmic rescoring. The acoustic model produces a decoding lattice, which is refined using a \textbf{}-Independent Static--Dynamic Rhythmic Model (TI-SDRM) that integrates long-term rhythmic structure with short-term adaptive dynamics through an adaptive interpolation mechanism. We curate a new real-world tabla solo dataset and a complementary synthetic dataset, establishing the first benchmark for weakly supervised TST in Hindustani classical music. Experiments demonstrate consistent and substantial reductions in stroke error rate over acoustic-only decoding, confirming the importance of explicit rhythmic structure for accurate transcription.
Paper Structure (52 sections, 19 equations, 1 figure, 7 tables, 1 algorithm)

This paper contains 52 sections, 19 equations, 1 figure, 7 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Rhythmic Structure in Hindustani classical music and Tabla
  4. Tabla Stroke Transcription (TST)
  5. Weakly Supervised Sequence Learning in ASR
  6. Music Language Modelling for Transcription
  7. Proposed Method
  8. Problem Description and Acoustic Model
  9. $T\bar{a}la$-Independent Static--Dynamic Rhythmic Model (TI-SDRM)
  10. $T\bar{a}la$-Independent Static Rhythmic Prior and Posterior Estimation
  11. Estimation of the $t\bar{a}la$ Posterior
  12. Dynamic Bayesian Model with Temporal Forgetting
  13. Temporal Adaptation with Exponential Forgetting.
  14. Initialization of Dirichlet Parameters
  15. Adaptive Interpolation Weight $\lambda_k$
  16. ...and 37 more sections

Figures (1)

  • Figure 1: Graphical illustration of state expansion for lattice-based rhythmic rescoring. An acoustic lattice $\mathcal{L}=(V,E,w_{\mathrm{ac}})$ obtained from CTC decoding merges multiple stroke histories at a lattice node $v$. During rhythmic rescoring, each history is expanded into a separate decoding state $(v,\mathbf{h}_i,\alpha_i)$ to preserve history-dependent rhythmic context. Candidate next strokes are scored using the combined acoustic--rhythmic objective (Eq. \ref{['eq:score_e']}), expanded states are updated and pruned using beam search (Eq. \ref{['eq:beam_p']}), and Viterbi decoding over the expanded lattice $\mathcal{L}_{\mathrm{out}}$ selects the single highest-scoring path as the final transcription $\hat{\mathbf{s}}$.