Contribution of spectral shapes to tone perception

Natthawut Kertkeidkachorn, Surapol Vorapatratorn, Sirinart Tangruamsub, Proadpran Punyabukkana, Atiwong SuchatoInterspeech 2012

TL;DR

This paper investigates how spectral shapes, beyond fundamental frequency (F0), contribute to the perception of Thai tones. The study reveals that while F0 is crucial, spectral information, particularly around the first formant, plays a significant role in accurate tone identification, especially in longer utterances.

What Problem Does This Paper Address?

Tones are fundamental to meaning in tonal languages like Thai. Traditionally, tone perception and recognition have focused almost exclusively on the fundamental frequency (F0) of speech, specifically its contour over time. However, human listeners are remarkably adept at distinguishing tones, often outperforming machine systems that rely solely on F0. This discrepancy suggests that other acoustic cues might be at play. This paper aims to explore these additional cues, with a particular focus on spectral shape, to understand their contribution to tone perception in Thai.

Methods

The researchers conducted a series of perception experiments using carefully crafted speech stimuli. They started with an "original stimulus set" of mono-, bi-, and tri-syllabic Thai utterances. Then, they created several "modified stimulus sets" by manipulating specific acoustic features:

  • Fo set: Sinusoidal signals with F0 contours matching the original stimuli, but with constant amplitude.
  • Fo+E set: Sinusoidal signals with F0 contours and amplitudes modulated by the original stimuli's energy envelope.
  • Lo_Fl set: Original stimuli low-pass filtered at a frequency related to the first formant.
  • Lo_F1/F2 set: Original stimuli low-pass filtered at a frequency between the first and second formants.
  • Lo_F2/F3 set: Original stimuli low-pass filtered at a frequency between the second and third formants.

Native Thai speakers listened to these stimuli and identified the tones of each syllable. Statistical tests (McNemar's tests) were used to compare the accuracy of tone identification across different stimulus sets.

Datasets

The study used a custom dataset of recorded Thai speech. The utterances consisted of the syllable /l a:/ in mono-, bi-, and tri-syllabic forms, with various tone combinations. These recordings formed the "original stimulus set," which was then processed to create the modified stimulus sets for the experiments.

Key Insights and Ideas

  • F0 is not enough: While F0 contours are the primary definition of Thai tones, they alone are insufficient for perfect human tone identification, especially in continuous speech.
  • Spectral shape matters: The study strongly suggests that spectral shape, particularly information within the vicinity of the first formant, contributes significantly to tone perception.
  • Syllable count amplifies spectral importance: The contribution of spectral shape to tone perception appears to increase with the number of syllables in an utterance.
  • Energy envelope is less critical for Thai tones: Unlike some other languages, the temporal changes in syllable energy did not show a significant impact on Thai tone perception in this study.

Key Results

  • Experiment #1 (Fo, Fo+E): The original stimuli yielded the highest accuracy. The Fo+E and Fo sets showed similar performance, suggesting energy modulation had little additional benefit over F0 alone. A significant gap between original and Fo-only stimuli highlighted the importance of other cues.
  • Experiment #2 (Lo_Fl, Lo_F1/F2, Lo_F2/F3): The Lo_Fl set (low-pass filtered around the first formant) resulted in the lowest accuracy, indicating that removing spectral information below the first formant significantly hinders tone perception. Conversely, filtering at higher frequencies (Lo_F1/F2, Lo_F2/F3) did not degrade performance as much, and in some cases, even slightly improved it compared to Lo_Fl.
  • Confusion Matrices: The most frequent confusion was between the low tone and the rising tone, which share similar F0 contours. This further supports the idea that spectral cues might help differentiate these tones.

Limitations and Future Work

  • Limited Stimuli: The study used only one syllable (/l a:/) and one male speaker. This limits the generalizability of the findings to other phonetic contexts or voices.
  • Multilingual Speakers: While the participants were native Thai speakers, the paper mentions that some previous studies (cited in the literature review) used multilingual but non-native Thai speakers, which could introduce different perceptual strategies.
  • Future Work: The authors encourage further research into the specific spectral features that contribute to tone perception and how these might interact with other acoustic cues in different tonal languages. Investigating the role of the first formant more deeply is a promising avenue.

Summary updated

BibTeX
@inproceedings{kertkeidkachorn2012contribution,
  author = {Natthawut Kertkeidkachorn and Surapol Vorapatratorn and Sirinart Tangruamsub and Proadpran Punyabukkana and Atiwong Suchato},
  title = {{Contribution of spectral shapes to tone perception}},
  booktitle = {Interspeech},
  year = {2012},
}
Figure