Modeling Formant Dynamics in Mandarin /ai/: Effects of Speech Style and Speech Rate

Yang Xiang, Jingyi Sun2025 · DOI

TL;DR

This study investigates how speech rate and style (conversational vs. read speech) influence the acoustic realization of the Mandarin diphthong /ai/. Using advanced statistical modeling, the researchers found that while both faster speech and more "clear" speech styles lead to higher formant values, they represent distinct dimensions of vocalic variation, suggesting speakers can independently adjust clarity and speed.

What Problem Does This Paper Address?

Understanding how speech rate affects vowel production is a long-standing question in phonetics. The "target undershoot" model suggests that in faster speech, articulators don't reach their full targets, leading to vowel reduction. However, previous studies have yielded conflicting results, with some finding no consistent pattern of vowel reduction across different speech rates. Another related concept is "speech clarity," often explained by the Hyper- and Hypo-articulation (H&H) theory, which posits that speakers adjust their articulation based on listener needs. This paper aims to disentangle the effects of speech rate and speech style (as a proxy for hyper-articulation) on the dynamic acoustic properties of the Mandarin diphthong /ai/. Specifically, it questions whether speech clarity and speech rate are interchangeable factors or distinct dimensions of vocalic variation.

Methods

The researchers employed Generalized Additive Mixed Models (GAMMs) to analyze the complex, non-linear relationships between vowel duration, normalized time within the diphthong, and formant frequencies (F1 and F2) across different speech styles. GAMMs are well-suited for this type of time-series data, capturing autocorrelations and non-linear patterns. They extracted formant values from audio recordings using auto-correlation and refined formant trajectories using a Viterbi-based dynamic tracking algorithm. To assess speech style, they compared read speech (assumed to be more hyper-articulated) with conversational speech. They also calculated spectral transition rates to quantify the speed of formant changes within the diphthong.

Datasets

The study utilized a Mandarin corpus comprising 20 hours of conversational speech and 6 hours of read speech from 30 native speakers. This corpus was previously published and included recordings of both spontaneous dialogues and read texts. The researchers focused on male tokens for their analysis, resulting in a dataset of 8,948 tokens.

Key Insights and Ideas

  • Speech Rate and Style are Distinct: The findings strongly suggest that speech rate and speech style are not interchangeable. While both faster speech and read speech (hyper-articulated) lead to higher F1 onset and F2 offset for /ai/, read speech exhibits a broader formant frequency range and maintains clearer articulation even at extreme durations. This implies speakers can independently control clarity and speed.
  • Vowel Duration's Complex Influence: Vowel duration has a significant, non-linear impact on formant trajectories. F1 onset generally increases with duration, while F1 offset initially decreases and then increases. F2 offset also increases with duration, suggesting that longer vowels allow for more complete articulation.
  • Shorter Vowels Require More Effort: Shorter diphthongs exhibit steeper spectral transition rates. This suggests that when time is limited, speakers exert more effort to maintain the diphthong's distinct acoustic targets. Conversely, longer vowels allow for more relaxed articulation, leading to slower transitions.
  • Read Speech Maintains Targets: Read speech, even at very fast or very slow rates, tends to maintain distinct acoustic targets compared to conversational speech. This supports the idea that read speech is produced with a higher degree of articulatory precision.

Key Results

  • Read speech showed an overall increase in F1 by approximately 26 Hz and F2 by nearly 30 Hz compared to conversational speech.
  • The study found significant style-specific temporal effects on both F1 and F2, indicating distinct articulatory dynamics between the two speech styles.
  • Transition rates generally decreased as vowel duration increased, contradicting the initial hypothesis that transition rates remain stable across different speech rates.
  • The study confirmed that read speech productions of /ai/ exhibited higher F1 onset and F2 offset, aligning with previous findings on clear speech.

Limitations and Future Work

The authors acknowledge that it is difficult to completely tease apart the effects of speech rate and speech style, as they are often intertwined. They also note that the corpus included ultrashort vowels (below 55 ms), which presented measurement challenges. While they used linear interpolation and safeguards to mitigate these issues, it remains a limitation. Future work could explore these dynamics in other Mandarin vowels and diphthongs, as well as investigate the role of other factors like following consonants and lexical tone in shaping vowel duration and formant trajectories. The study also suggests that further research could directly manipulate speech clarity to confirm its independent effects.

Summary updated

BibTeX
@misc{xiang2025modeling,
  author = {Yang Xiang and Jingyi Sun},
  title = {{Modeling Formant Dynamics in Mandarin /ai/: Effects of Speech Style and Speech Rate}},
  year = {2025},
  doi = {10.21437/interspeech.2025-968},
}

Loading related content...

Figure