Automatic labeling of tabla signals

Olivier Gillet, Gaël RichardISMIR 2003

Introduction

Automatic transcription of musical signals is a cornerstone of Music Information Retrieval (MIR). While much research has focused on Western pitched instruments, percussive instruments like the Tabla—a pair of hand drums central to North Indian classical music—present unique challenges. This paper presents a robust system for the automatic transcription of Tabla sequences into their constituent mnemonic syllables, known as bols.

The Challenge of Tabla Transcription

Transcribing Tabla music is complex due to two primary factors:

  1. Acoustic Overlap: The resonant nature of the drums means that the decay of one stroke often overlaps with the onset of the next, altering its timbre.
  2. Context-Dependent Notation: Bols (e.g., Dha, Tin, Ge) are not just sounds but represent specific fingerings. Different bols can sound acoustically similar but appear in distinct rhythmic contexts. Consequently, isolated stroke classification is insufficient; the sequence must be modeled as a whole.

Methodology

The proposed system adopts a pipeline inspired by large-vocabulary continuous speech recognition, treating rhythmic phrases as "sentences" composed of bol "words."

Figure 1: Block diagram of the automatic tabla transcription and synthesis system.

Segmentation and Feature Extraction

The audio stream is first segmented using an envelope-and-threshold onset detection method. For each detected stroke, the system extracts a 12-dimensional feature vector. This is achieved by modeling the power spectrum as a weighted sum of four Gaussian distributions across specific frequency bands. This compact representation captures the essential timbral characteristics of different strokes.

Figure 2: Normalized amplitude spectra for various tabla bols (Dha, Ge, Na, Ti, Ke), illustrating the spectral diversity used for feature extraction.

Sequence Modeling with HMMs

The core of the system is a Hidden Markov Model (HMM) that incorporates an n-gram language model (3-gram or 4-gram). This allows the system to leverage the "grammar" of Tabla compositions—where certain sequences of bols are highly probable—to resolve acoustic ambiguities. The Viterbi algorithm is employed to decode the most likely sequence of bols from the extracted features.

Experimental Results

The system was evaluated on a custom corpus of 5,715 bols across 64 phrases, recorded using three different instruments and environments to test robustness.

  • Recognition Accuracy: The HMM-based approach achieved a peak recognition rate of 93.6% using 4-grams and 2 Gaussian mixtures.
  • Segmentation Performance: The onset detector correctly identified 98.3% of strokes, benefiting from the relatively sparse polyphony of Tabla music.
  • Generalization: When tested on an unseen instrument in a noisy environment, the system maintained high performance (88-90%), though simpler models (3-grams) proved more robust against overfitting in these scenarios.

Applications and Conclusion

The system operates in real-time and has been integrated into the Tablascope software. This tool enables the use of a Tabla as a MIDI controller and provides automated phrase correction based on traditional grammar. By treating percussion as a structured language, this research demonstrates that high-accuracy transcription is possible even for complex, non-pitched instruments. Future work will explore the application of this HMM-based language modeling to a broader range of percussive signals.

Summary updated

BibTeX
@inproceedings{gillet2003automatic,
  author = {Olivier Gillet and Gaël Richard},
  title = {{Automatic labeling of tabla signals}},
  booktitle = {ISMIR},
  year = {2003},
}
Figure