Tabla Stroke Transcription (TST)

• Tabla Stroke Transcription (TST) is the computational conversion of tabla recordings into discrete stroke symbols (e.g., Dha, Tin, Na), facilitating detailed rhythmic analysis.
• State-of-the-art methods use techniques like CRNN, MAML, and CTC on log-Mel spectrograms to extract and classify stroke features with high precision.
• Rhythm-aware sequence models incorporating tāla-informed priors enhance transcription accuracy despite challenges like timbral overlap and improvisational phrasing.

Tabla Stroke Transcription (TST) refers to the systematic computational process of converting raw audio recordings of tabla performances into discrete, time-ordered sequences of percussive "stroke symbols" (e.g., Dha, Tin, Na), capturing the intricate rhythmic language foundational to Hindustani classical music. This translation of acoustic signals into symbolic stroke notations supports structural rhythm analysis, musicological research, and pedagogical tools, but is challenged by the high timbral overlap between stroke classes, improvisational phrasing, and limited annotated data (Kodag et al., 13 Jan 2026).

1. Stroke Taxonomies and Dataset Labeling

Precise definition and granularity of stroke classes are crucial for both musical relevance and computational tractability. Multiple taxonomic approaches are evident in recent literature:

• Fine-Grained Bol Taxonomies: Datasets encapsulate as many as 18–30 distinct tabla bols, as in the tabla-solo dataset $D_{T1}$ ({Da, Ki, Ge, Ta, Na, Din, Kda, Tit, Dha, ...}) and the Hindustani concert dataset $D_{T3}$ ({Dha, Dhin, Tin, Na, Tun, Kat, ...}) (Kodag et al., 2024).
• Acoustic/Musicological Groupings: A reduced taxonomy into four robust classes is suited for accompaniment analysis: Damped, Resonant Treble, Resonant Bass, Resonant Both (A. et al., 2021).

Data labeling varies:

• Strong Supervision: Frame-wise one-hot labeling with explicit stroke onsets (Kodag et al., 2024, Kodag et al., 8 Jan 2025).
• Weak Supervision: Ordered, unaligned symbolic stroke sequences without temporal annotation, addressing annotation cost and scaling (Kodag et al., 13 Jan 2026).

Datasets span synthetic (constructed from isolated strokes), solo tabla, and polyphonic concert recordings. Support sets for meta-learning often consist of as few as 32 examples per stroke class.

2. Signal Processing and Feature Extraction

All state-of-the-art TST systems transform audio into log-Mel spectrograms ($X\in\mathbb{R}^{F\times T}$, e.g., $F=128$ Mel bands, window length $46.4$ ms, $10$ ms hop) with per-feature normalization. Feature representations range from full spectrograms (CRNN-based), to segment-level descriptors engineered for timbral and temporal discrimination (Kodag et al., 2024, Kodag et al., 13 Jan 2026, A. et al., 2021).

• Handcrafted Features (four-class taxonomy) (A. et al., 2021):
  • Spectral centroid, skewness, kurtosis.
  • 13 MFCCs.
  • Spectral flux, short-time energy.
  • Log attack time, temporal centroid, zero-crossing rate.
  • Band-specific decay envelope features (50–200 Hz, 200–2000 Hz): decay rates, intercepts, spline knots, $R^2$.
  • Energy deltas with respect to previous stroke.
• Deep Feature Learning (CRNN/TDNN) (Kodag et al., 2024, Kodag et al., 8 Jan 2025, Kodag et al., 13 Jan 2026):
  • Three-layer convolutional front-ends extract time–frequency patterns.
  • Bidirectional RNNs (e.g., Bi-GRU) or C-TDNN-F networks model stroke temporal structure and context.
  • Frame-wise classification via softmax over $C$+1 stroke/no-stroke classes.

3. Learning Paradigms: Supervised, Meta-learning, and Weakly-supervised

Supervised learning approaches require detailed onset annotations; labeled data scarcity limits scalability and generalization (Kodag et al., 2024, Kodag et al., 8 Jan 2025). Recent advances shift toward:

• Model-Agnostic Meta-Learning (MAML) (Kodag et al., 2024, Kodag et al., 8 Jan 2025):
  • The model parameters are structured as $[\theta_1, \phi]$—with frozen convolutional layers $\theta_1$ and adaptable recurrent/classifier parameters $\phi$.
  • For each meta-task $T_i$, perform $N$ inner-loop gradient steps on a support set $T^s_i$, yielding adapted parameters $\phi^i_N$,

  $$\phi_n^i = \phi_{n-1}^i - \alpha\nabla_\phi L_{T_i^s}(f_{[\theta_1, \phi_{n-1}^i]})\,, \quad n=1,\ldots,N$$ - The global meta-objective optimizes post-adaptation loss on task query sets:

  $$\phi \leftarrow \phi - \beta \sum_{T_i} \nabla_\phi L_{T_i^q}(f_{[\theta_1, \phi_N^i]})$$ - Enables rapid few-shot adaptation to new stroke inventories with as little as $s=32$ examples per class.

• Weakly-supervised CTC with Rhythmic Rescoring (Kodag et al., 13 Jan 2026):

  • Trains an acoustic model on unaligned symbolic stroke sequences via Connectionist Temporal Classification (CTC):

  $$L_{CTC}(X,Y) = -\log\left( \sum_{\pi\in\mathcal{B}^{-1}(Y)} \prod_{t=1}^T p_\theta(\pi_t \mid X) \right)$$ - Candidate transcriptions are decoded as a lattice and refined with a structured rhythmic model (TI-SDRM) that fuses long-span tāla priors and adaptable local dynamics via adaptive Jensen-Shannon-divergence-weighted interpolation.

Handcrafted, feature-based classification remains relevant for small datasets and reduced taxonomies, utilizing Random Forests on spectral, temporal, and decay features, with data balancing via oversampling, SMOTE, and pitch-shifting augmentations (A. et al., 2021).

4. Rhythmic Modeling and Sequence-Level Correction

Purely acoustic models struggle with ambiguities arising from tabla’s high timbral overlap and improvisational phrasing. Sequence-level models addressing rhythm are employed at two granularities:

• Explicit Tāla-Informed Priors (Kodag et al., 13 Jan 2026):

  • Static long-term prior $P_s$, marginalized over possible tāla cycles, computed as an $n$-gram model over stroke sequences.
  • Online dynamic short-term model $P_d$ adapts to local tempo and improvisation via Dirichlet-multinomial transitions with exponential forgetting.
  • Final probability: $$P_{\text{comb}}(s_k|h) = (1-\lambda_k)P_s(s_k|h) + \lambda_k P_d(s_k|h)$$ with data-driven interpolation.
• Sequence Alignment for Tāla Identification (Post-transcription) (Kodag et al., 2024, Kodag et al., 8 Jan 2025):
  • Needleman–Wunsch sequence alignment between transcription and thekā reference.
  • Stroke-ratio (cosine similarity) between the occurrence vector of transcribed and reference strokes.
  • These measures enable robust $t\bar{a}la$ identification, even under noisy or incomplete transcriptions.

A plausible implication is that the integration of context-sensitive sequence models—especially those respecting indigenous metrical structures—is crucial for overcoming the limitations of frame- or segment-level prediction, particularly in low-resource or polyphonic scenarios.

5. Evaluation Protocols and Empirical Results

Evaluation employs dataset-specific splits, onset-error tolerances (e.g., 50 ms collar), and class-balanced metrics:

• Frame-wise F1-score for onset and stroke label at frame-level (Kodag et al., 2024, Kodag et al., 8 Jan 2025).
• Stroke Error Rate (SER): $(S + I + D)/N$, analogous to Word Error Rate, at symbolic sequence level (Kodag et al., 13 Jan 2026).
• Balanced Accuracy and F1 per stroke class, critical under class imbalance (A. et al., 2021).

Performance, under meta-learning and weak supervision, substantially improves over transfer learning and frame-level-only baselines:

Model/Setting	In-domain F1 (%)	Cross-domain F1 (%)	Weak-Sup. SER (%)
CRNN (PTM1), full seq.	93.2	–	–
PTM1+MAML (meta-test)	81.3	63.0	–
PTM1+Transfer (meta-test)	62.2	39.5	–
CTC Acoustic only (DT1)	–	–	38.6
+TI-SDRM rescoring (DT1)	–	–	30.9
Four-class RF baseline (test)	59	–	–
Four-class RF + pitch shift	65	–	–

Only 3 meta-learning adaptation steps on 32 examples yield $>80$\% F1 in-domain, $>60$\% F1 cross-domain (Kodag et al., 2024, Kodag et al., 8 Jan 2025). TI-SDRM delivers 20–37% absolute reductions in Stroke Error Rate relative to acoustic-only CTC (Kodag et al., 13 Jan 2026).

6. Limitations and Future Directions

Major challenges persist:

• Strong Supervision Cost: Onset-level annotation is resource-intensive; weak supervision mitigates the bottleneck by relying on symbolic label sequences (Kodag et al., 13 Jan 2026).
• Timbre and Instrument Variability: Key features shift across instruments and tuning; pitch-shift augmentation partially addresses this, but more structured augmentation (e.g., simulating decay envelopes, spectrotemporal modifications) is still undeveloped (A. et al., 2021).
• Context Ignorance in Segment-based Models: Current Random Forest approaches lack the ability to exploit thekā or metrical context. Sequence models, either as n-grams or neural LMs, are promising for disambiguation (A. et al., 2021, Kodag et al., 13 Jan 2026).
• Source Separation: TST from polyphonic mixtures (vocals, melodic instruments) remains challenging, despite the relative robustness of meta-learned models compared to baseline transfer learning (Kodag et al., 8 Jan 2025). Source separation integration is anticipated as a necessary component for concert audio.
• Generalization: While meta-learning and weakly supervised approaches enable few-shot and low-resource adaptation, further work is needed on cross-instrument and cross-corpus robustness.

A plausible implication is that future TST research will increasingly integrate rhythm-aware sequence models, weakly supervised learning on large unlabeled corpora, and transferable acoustic representations aligned with musicological structure.

7. Applications and Impact

TST underpins diverse applications:

• Musicological Analysis: Enables cycle detection, metric analysis, and studies of improvisation and expressivity in Indian music (Kodag et al., 13 Jan 2026).
• Archival and Retrieval: Supports large-scale corpus annotation, search, and indexing by rhythmic motifs (Kodag et al., 2024).
• Pedagogy and Practice: Facilitates automatic feedback for students, digital companions for learning, and rhythm pattern generation and identification (Kodag et al., 2024).
• General Percussion Transcription: CRNN+MAML pipelines generalize to low-resource drum transcription beyond tabla, with superior gains over state-of-the-art on multiple Western drum datasets (Kodag et al., 8 Jan 2025).

Recent releases of weakly labeled datasets and modern TST systems establish both methodological and benchmarking baselines, accelerating progress in automatic rhythm and percussion transcription across global traditions.

---

Key References:

• $T\bar{a}laGen:$ A System for Automatic $T\bar{a}la$ Identification and Generation (Kodag et al., 2024)
• Meta-learning-based percussion transcription and $t\bar{a}la$ identification from low-resource audio (Kodag et al., 8 Jan 2025)
• Weakly Supervised Tabla Stroke Transcription via TI-SDRM: A Rhythm-Aware Lattice Rescoring Framework (Kodag et al., 13 Jan 2026)
• Automatic Stroke Classification of Tabla Accompaniment in Hindustani Vocal Concert Audio (A. et al., 2021)