Acoustic and prosodic analysis of Japanese vowel-vowel hiatus with laryngeal effect

Kitazawa Shigeyoshi, Shinya KiriyamaInterspeech 2004

TL;DR

This paper investigates how Japanese speakers manage vowel-vowel (V-V) hiatuses, which occur when adjacent phrases end and begin with the same vowel. The researchers found that speakers use subtle laryngeal effects like glottalization and nasalization, along with duration adjustments, to signal these boundaries, making them perceptible to listeners. This work contributes to a deeper understanding of prosodic boundaries in speech, crucial for improving speech recognition and synthesis systems.

What Problem Does This Paper Address?

In fluent speech, distinguishing between word and phrase boundaries can be challenging, especially when the transition involves similar sounds. This is particularly true in Japanese, a language with open syllables, where vowel-vowel (V-V) hiatuses are common. These hiatuses occur when the final vowel of one accentual phrase is the same as the initial vowel of the next. While linguistic theories suggest mechanisms for marking these boundaries, the precise acoustic and prosodic cues used by speakers, and how listeners perceive them, remain an area for detailed investigation. This paper aims to shed light on these cues, focusing on the laryngeal effects and duration changes that help resolve V-V hiatuses in Japanese.

Methods

The study employed a multi-modal approach combining perceptual listening tests with acoustic analysis.

  • Perceptual Experiments: Shortened speech segments around V-V hiatuses were presented to listeners who rated their naturalness on a scale. These segments were systematically varied by shifting the boundary point by small increments (one vocal cord vibration period).
  • Acoustic Analysis:
    • Electroglottography (EGG): This technique measured vocal fold vibration, providing data on the open quotient (the proportion of the glottal cycle during which the vocal folds are open) and fundamental frequency (F0). Changes in these parameters can indicate laryngeal events like glottalization.
  • Spectrogram Analysis: Wide-band and narrow-band spectrograms were used to visualize speech sounds, helping to identify acoustic features like nasalization and spectral changes.
  • Waveform Analysis: Speech and EGG waveforms were analyzed to observe temporal patterns and F0 variations.
  • Prosodic Labeling: The J-ToBI (Japanese ToBI) prosody annotation scheme was used to label the speech data, though the authors note that J-ToBI doesn't explicitly describe V-V hiatuses.

Datasets

The primary dataset used was the Japanese MULTEXT prosody corpus. This corpus consists of speech data recorded by native speakers of the Tokyo dialect, reading texts designed to elicit simulated spontaneous utterances. The specific materials analyzed for V-V hiatuses included:

  • Particle-vowel transitions (36 examples)
  • Adjective-vowel transitions (5 examples)
  • Word-word transitions (4 examples)

The corpus was created based on the EUROM1 specification for multi-language prosody databases.

Key Insights and Ideas

  • Laryngeal Effects are Key: The most significant finding is that subtle laryngeal adjustments, particularly phrase-initial glottalization and phrase-final nasalization, play a crucial role in marking V-V hiatus boundaries.
  • Glottalization as a Primary Cue: Phrase-initial glottalization, characterized by a drop in F0 and a dip in the EGG open quotient, was observed in a majority of the analyzed hiatuses. This suggests it's a primary acoustic cue for signaling the start of a new phrase.
  • Nasalization as a Complementary Cue: Phrase-final nasalization, particularly in particles like "ga," also contributes to boundary perception, especially when contrasted with a non-nasalized following vowel.
  • Duration Matters: Both phrase-final lengthening and phrase-initial shortening were observed. This temporal adjustment helps maintain the rhythmic structure of Japanese and aids in segmenting the V-V sequence.
  • Perception Aligns with Acoustics: Perceptual judgments of naturalness generally correlated with the presence and strength of these acoustic cues, suggesting listeners are sensitive to these subtle phonetic details.
  • Syntactic Context Aids Perception: While acoustic cues are important, the authors suggest that syntactic context likely plays a role in resolving hiatuses, and human perception might be more sensitive to these cues than current signal processing methods can detect.

Key Results

  • Out of 45 analyzed hiatuses, 17 showed clear glottalization, and 23 showed weak glottalization accompanied by other features.
  • 5 hiatuses showed phrase-final nasalization as the primary cue.
  • Perceptual experiments indicated that listeners could identify natural boundaries, and these points often coincided with acoustic evidence of glottalization and duration changes.
  • The study found a trend where perceptual scores increased with higher "break indices" (BI), suggesting that more disjunctive boundaries are perceived as more natural in certain contexts.
  • The ratio of the duration of the phrase-ending vowel to the phrase-initiating vowel averaged 1.7, indicating a tendency for the ending vowel to be longer.

Limitations and Future Work

  • Limited Dataset Size: The authors acknowledge that the number of samples (45 hiatuses) is too small to generalize the results broadly.
  • J-ToBI Limitations: The J-ToBI scheme, while useful, does not explicitly annotate V-V hiatuses, requiring the researchers to adapt its use.
  • Future Work: The authors suggest that further research should explore the sensitivity of human perception to these acoustic cues, especially in cases where acoustic features are subtle. Investigating how syntactic emphasis influences these boundary marking strategies is also a promising direction. The potential for these findings to improve speech recognition and synthesis systems is highlighted.

Summary updated

BibTeX
@inproceedings{shigeyoshi2004acoustic,
  author = {Kitazawa Shigeyoshi and Shinya Kiriyama},
  title = {{Acoustic and prosodic analysis of Japanese vowel-vowel hiatus with laryngeal effect}},
  booktitle = {Interspeech},
  year = {2004},
}
Figure