Acoustic-Prosodic Features of Tabla Bol Recitation and Correspondence with the Tabla Imitation

Rohit M A, Preeti RaoInterspeech 2018

TL;DR

This paper explores the expressive vocalizations (bols) used in North Indian classical tabla music, investigating how their acoustic prosody (pitch, intensity, duration) mirrors the actual tabla drumming. By analyzing a novel dataset of synchronized bol recitations and tabla performances, they found that vocal intensity variations strongly correlate with tabla intensity, while pitch correspondences are primarily limited to the pitch-varying bass drum.

What Problem Does This Paper Address?

In North Indian classical music, tabla players use spoken syllables called "bols" to represent drum strokes. This oral tradition is crucial for teaching and practice, with bol recitations being highly expressive, much like spoken language. While previous research has focused on matching individual bol sounds to drum strokes, this paper tackles a more nuanced problem: understanding how the supra-segmental prosodic features of vocal bol recitation (like pitch contours and intensity variations across phrases) correspond to the acoustic features of the actual tabla performance. This is challenging because the human voice and the tabla are fundamentally different sound production systems, and tabla scores often lack explicit dynamic markings. The authors aim to bridge this gap by investigating if the expressive nuances in vocal recitation are encoded in the tabla's sound, potentially leading to better models for synthesized tabla music.

Methods

The researchers employed a multi-stage analysis:

  • Dataset Creation: They utilized an existing dataset of tabla compositions and bol recitations but found the original audio quality challenging for detailed prosodic analysis due to accompanying harmonium. To address this, they arranged for a senior tabla disciple to re-record 20 compositions in a sound-proof studio, ensuring synchronized, high-quality audio for both bol recitation and tabla playing.
  • Onset Detection and Time-Alignment: They used spectral flux analysis to detect the precise onsets of syllables in the bol recitation and strokes in the tabla performance. These detected onsets were then time-aligned with the written musical scores to segment the audio accurately.
  • Feature Extraction: For both the vocal recitation and the tabla performance, they extracted key prosodic features:
  • Pitch (F0): Measured using Praat, with low-pass filtering applied to tabla audio to isolate the bass drum's pitch.
  • Intensity: Calculated as short-time energy, normalized, and expressed in decibels.
  • Duration: Implicitly captured through the segmentation of bols/strokes. These features were normalized and represented on a logarithmic scale for analysis.

  • Correlation Analysis: They performed two levels of correlation:

  • Cycle-level: Examining the correlation of mean intensity values across entire 16-beat cycles of the tabla rhythm.
  • Bol/Stroke-level: Analyzing the correlation of mean pitch and intensity for individual pitched bols and strokes, categorizing them into "Ghe-based" (involving both drums) and "pitched dayan" (treble drum).

Datasets

The primary dataset used is a newly recorded parallel corpus of 20 North Indian classical tabla compositions. This dataset consists of synchronized audio recordings of:

  1. Bol Recitation: The vocal performance of the tabla syllables.
  2. Tabla Performance: The actual drumming of the compositions.

This dataset was derived from an initial set of 38 compositions from Pt Arvind Mulgaonkar's "Shades of Tabla 1" DVD, but re-recorded to ensure higher audio quality and better synchronization for acoustic analysis. The re-recordings were performed with specific tuning for the dayan (treble drum, ~257 Hz) and bayan (bass drum, ~80 Hz).

Key Insights and Ideas

  • Vocal Prosody as a Performance Blueprint: The study highlights that vocal bol recitation is not just a mnemonic device but a rich carrier of expressive information, akin to spoken language prosody.
  • Intensity as a Shared Language: The most robust correspondence found was between the intensity variations in vocal recitation and the overall intensity of the tabla strokes. This suggests that the dynamic shaping of vocal delivery directly translates to the force and emphasis applied to the drums.
  • Pitch's Selective Role: While vocal pitch varies significantly, its direct correspondence to tabla sound is limited. This is primarily observed in the pitch-varying bass drum (bayan), where vocal pitch contours can influence the drum's pitch. The treble drum (dayan), with its fixed pitch, shows little correlation in this regard.
  • The "Ghe" Stroke's Complexity: Strokes involving both drums ("Ghe-based") showed better pitch correlation, likely due to the bayan's inherent pitch variability, which can be influenced by vocal cues.
  • Beyond Segmental Matching: The paper moves beyond simple sound-to-stroke mapping to explore the more complex, supra-segmental layer of musical expression.

Key Results

  • Cycle-level Intensity Correlation: A moderately strong positive correlation (Pearson's 'r' around 0.5) was observed between the mean intensity of vocal recitation and tabla playing across rhythmic cycles.
  • Bol/Stroke-level Correlations:
    • Ghe-based strokes: Mean pitch (F0) showed a significant correlation (r=0.37, p<0.05) between recitation and playing. Mean intensity also correlated well (r=0.39, p<0.05).
  • Pitched Dayan strokes: Mean pitch showed a weak correlation (r=0.085, p>0.05), while mean intensity showed a moderate correlation (r=0.35, p<0.05).
  • Vocal Prosody Characteristics: The syllable rate of recitation varied widely (4-14 bols/s), and the intensity and F0 ranges were comparable to lively conversational speech.

Limitations and Future Work

  • Limitations:
    • The presence of a harmonium lehera (melodic cycle marker) in the original recordings made detailed prosodic analysis difficult, necessitating re-recording.
  • While the dataset was re-recorded, the accompanying harmonium in the new recordings still posed challenges for precise pitch and intensity tracking.
  • The study focused on Tintal, a specific 16-beat cycle, and its findings might not generalize to other talas.
  • The analysis of pitch for the bayan was limited to a low-pass filtered signal, which might not capture all nuances.
  • Future Work:
    • Enhance audio processing techniques for better separation of overlapping strokes in time and frequency, leading to more accurate analyses.
  • Manually annotate the dataset for rhythmic accents and phrase boundaries to explore other musicological functions of prosodic variation in oral recitation.
  • Investigate the role of prosody in demarcating rhythmic phrase boundaries.
  • Develop models for synthesized tabla music that incorporate these supra-segmental prosodic features to increase naturalness.

Summary updated

BibTeX
@inproceedings{a2018acoustic,
  author = {Rohit M A and Preeti Rao},
  title = {{Acoustic-Prosodic Features of Tabla Bol Recitation and Correspondence with the Tabla Imitation}},
  booktitle = {Interspeech},
  year = {2018},
}
Figure