TL;DR
This paper investigates the prevalence and impact of laryngealization (creaky voice) in Mandarin and Cantonese, arguing that its common occurrence, especially in low-tone syllables, significantly disrupts standard pitch-based tone recognition. The authors propose that laryngealization itself could be a valuable feature for improving tonal recognition accuracy.
What Problem Does This Paper Address?
Recognizing tones in tonal languages like Mandarin and Cantonese is crucial for speech technology. Current systems primarily rely on fundamental frequency (f0) contours. However, a phenomenon called laryngealization, or creaky voice, often co-occurs with low tones in these languages. This irregular phonation type makes f0 difficult to accurately measure, potentially leading to errors in tone recognition. While laryngealization is known to affect perception, its systematic impact on acoustic feature extraction for automatic tone recognition and its potential as a feature itself have not been thoroughly explored. This paper aims to fill that gap by quantifying the prevalence of laryngealization and discussing its implications for tonal recognition systems.
Methods
The researchers recorded controlled speech corpora of Mandarin and Cantonese, designed to elicit specific tonal contexts. They then analyzed these recordings using Praat, employing a multi-faceted approach to label laryngealization. This involved examining:
- Waveform irregularities: Specifically looking for irregular glottal pulses (vocal fry) or period doubling.
- Spectrogram analysis: Identifying loss of definition in harmonic structure in narrow-band spectrograms.
- f0 detection failure: Noting instances where standard f0 extraction algorithms (like Praat's autocorrelation) failed.
- Auditory perception: Relying on the subjective percept of laryngealization/creaky voice quality.
These criteria were applied consistently to identify laryngealized vs. non-laryngealized syllables. The study then analyzed the distribution of laryngealization across different tones and syllables, and discussed its implications for f0 feature extraction and potential use as a recognition feature.
Datasets
The study utilized newly recorded, controlled speech corpora for:
- Cantonese: 12 native speakers from Hong Kong and Macau, producing minimal sets of bitones over 6 tones, with 5 repetitions each. The analysis focused on sentence-medial bitones.
- Mandarin: 12 native speakers from Beijing, producing minimal sets of bitones over 5 tones (including neutral tone), uttered in sentence-initial, medial, and final positions. The analysis focused on sentence-medial bitones, excluding Tone 3 realized as Tone 2 due to sandhi.
The total number of analyzed examples was 1440 for Cantonese and 1280 for Mandarin. These corpora were designed to control for speech rate (around 3.2 syllables/second) and contextual tonal variability.
Key Insights and Ideas
- Laryngealization is prevalent and disruptive: Contrary to the assumption that f0-based systems are robust, laryngealization is common in low-tone syllables of Mandarin (Tone 3) and Cantonese (Tone 4), and significantly interferes with accurate f0 extraction.
- Laryngealization is a perceptual cue: Previous studies show listeners use laryngealization to distinguish tones, particularly in Cantonese Tone 4, suggesting it's not just noise but a meaningful acoustic signal.
- Current f0 processing might mask the problem: The paper speculates that manual correction, smoothing, and averaging of f0 contours in existing tone recognition systems might be implicitly filtering out the effects of laryngealization, thus hiding the problem and the potential of laryngealization-based features.
- Laryngealization could be a feature, not just noise: Given its perceptual relevance and prevalence, laryngealization itself could be a powerful feature for improving tonal recognition, especially for tones that are otherwise confusable.
Key Results
- Cantonese: Laryngealization occurred on Tone 4 about a quarter of the time (25% overall), with a strong preference for the second syllable of bitones. For some speakers, over half of Tone 4 instances were laryngealized. Laryngealization was rare on other tones (2% on Tone 2).
- Mandarin: Laryngealization was more prevalent than in Cantonese, occurring on Tone 3 in 68% of instances across speakers. It also appeared on Tones 2 and 4 (8% and 5% respectively).
- Period Doubling: This specific type of laryngealization was observed more frequently in female speakers for both languages.
- f0 Extraction Issues: Examples show that laryngealization can lead to complete f0 detection failure over entire vowel segments, making standard f0-based features unreliable.
Limitations and Future Work
- Limitations:
- The study acknowledges that while laryngealization is prevalent, its impact on f0 feature extraction might be mitigated by current system practices (manual correction, averaging), which could be masking the true extent of the problem and the potential of laryngealization-based features.
- Previous work on voice quality features (e.g., [14]) showed limited success, possibly due to the features used or the reliance on f0 for their calculation.
- Future Work:
- The authors strongly suggest that laryngealization-based features should be considered for tonal recognition systems.
- Further research into automatic detection of irregular phonation (e.g., [16, 17, 18]) could provide useful alternatives for feature extraction.
- Investigating how different parametrizations of voice quality features might perform better than those previously studied is recommended.
- Addressing the disturbance of f0 tracking due to laryngealization is crucial, even if laryngealization-based features are not ultimately used.
Summary updated