The Contribution of Lip Protrusion to Anglo-English /r/: Evidence from Hyper- and Non-Hyperarticulated Speech

Hannah King, Emmanuel FerragneInterspeech 2019

TL;DR

This paper investigates the role of lip protrusion in producing the English /r/ sound, particularly in Anglo-English. By analyzing articulatory data from speakers producing /r/ in both normal and "hyperarticulated" (emphasized) speech, the researchers found that lip protrusion is a significant strategy, especially for speakers with a "bunched" tongue shape, and that this protrusion increases in hyperarticulated speech.

What Problem Does This Paper Address?

The production of the English /r/ sound is notoriously complex and variable. While previous research has focused on tongue shapes (like retroflex or bunched) in rhotic English varieties, less is known about how non-rhotic varieties like Anglo-English produce /r/, and specifically, how lip movements contribute. It's generally accepted that lip protrusion can accompany /r/, but the precise relationship between tongue shape, lip protrusion, and the acoustic outcome (specifically, the third formant, F3, which is crucial for distinguishing /r/) has remained unclear. This paper aims to fill this gap by examining the interplay between tongue articulation, lip protrusion, and acoustic properties in Anglo-English /r/, especially under conditions of hyperarticulation, where speakers might exaggerate their speech for clarity.

Methods

The study employed a multi-modal approach to capture articulatory and acoustic data.

  • Ultrasound Tongue Imaging: This technique provided real-time visualization of tongue movements, allowing researchers to classify /r/ into five types: Mid Bunched, Front Bunched, Front Up, Tip Up, and Curled Up.
  • Lip Videos: High-speed cameras captured both front and profile views of the lips to measure protrusion.
  • Acoustic Recording: Audio data was recorded to analyze formants, particularly F3.
  • Hyperarticulation Elicitation: A clever "silent speech" recognition task was used. Participants were told a computer was trying to recognize their speech from visual cues alone. When the computer "misidentified" words, participants were prompted to repeat them, leading to hyperarticulated speech. This allowed for a comparison between normal (non-hyperarticulated) and emphasized productions.
  • Data Analysis:
    • Statistical Analysis: Linear mixed-effects (LME) models were used to analyze the relationships between /r/ type, lip protrusion, vowel context, and hyperarticulation.
  • Deep Neural Networks (DNNs): ResNet-18 was trained on ultrasound and lip images to classify /r/ types and predict hyperarticulation. Class Activation Maps (CAMs) were used to visualize which parts of the images the DNNs focused on, helping to interpret their decisions.

Datasets

The study used data collected from 24 native Anglo-English speakers. The recordings included:

  • Nine /r/-initial monosyllabic words followed by various vowels.
  • Filler words starting with /w/.
  • Both non-hyperarticulated and hyperarticulated productions of these stimuli.
  • No pre-existing benchmark datasets were used; the data was collected specifically for this study.

Key Insights and Ideas

  • Lip Protrusion as a Key Strategy: The study strongly suggests that lip protrusion is not just a secondary effect but a significant articulatory strategy for producing English /r/, particularly in Anglo-English.
  • Bunched /r/ and Protrusion: Speakers who use a "bunched" tongue shape (tip down) tend to exhibit more lip protrusion than those using a "retroflex" shape (tip up). This aligns with the hypothesis that bunched /r/ speakers might use lip protrusion to compensate for a lack of sublingual space, a feature more prominent in retroflex /r/.
  • Hyperarticulation Enhances Protrusion: When speakers are prompted to hyperarticulate their speech (e.g., during error correction), lip protrusion significantly increases. This suggests speakers actively recruit more articulatory effort, including lip rounding, to enhance the distinctiveness of the /r/ sound.
  • Articulatory-Acoustic Trading Relations: The findings support the idea of "trading relations" where different articulatory gestures can achieve similar acoustic goals. Here, lip protrusion and tongue shape variations appear to work together to lower the third formant (F3), a key acoustic cue for /r/.
  • DNNs as a Tool for Interpretation: While DNNs achieved high classification accuracy, the use of CAMs revealed that models sometimes relied on non-phonetic cues (like hardware artifacts). This highlights the importance of interpretability methods when using DNNs in linguistic research.

Key Results

  • Lip Protrusion and /r/ Type: A statistically significant finding showed that increased lip protrusion leads to a lower probability of retroflexion (i.e., favors bunched configurations).
  • Hyperarticulation and /r/ Type: Hyperarticulated speech was associated with a significantly higher rate of retroflexion.
  • Hyperarticulation and Acoustics: Hyperarticulated /r/ exhibited more lip protrusion and lower F3 values, consistent with the hypothesis that speakers exaggerate articulatory features to enhance acoustic distinctiveness.
  • DNN Performance:
    • Classifying /r/ type using DNNs achieved mean test accuracies around 97-98% for ultrasound and front/profile lip images.
  • Predicting hyperarticulation using DNNs yielded accuracies of 77-88% for ultrasound and front/profile lip images.

Limitations and Future Work

  • Limitations:
    • The DNN analysis, while promising, showed that models sometimes relied on non-phonetic cues, indicating a need for careful interpretation and potentially more robust data augmentation or model architectures.
  • The study acknowledges that while hyperarticulation might aim for acoustic enhancement, it could also be driven by the need for visual intelligibility in the "silent speech" paradigm.
  • Future Work:
    • Investigate whether there are perceptible differences in the auditory and visual domains between retroflex and bunched /r/ with varying degrees of protrusion.
  • Train DNN models on whole videos rather than selected frames for potentially richer analysis.
  • Further explore the interplay between lip protrusion and tongue configuration in achieving acoustic goals for /r/.

Summary updated

BibTeX
@inproceedings{king2019contribution,
  author = {Hannah King and Emmanuel Ferragne},
  title = {{The Contribution of Lip Protrusion to Anglo-English /r/: Evidence from Hyper- and Non-Hyperarticulated Speech}},
  booktitle = {Interspeech},
  year = {2019},
}
Figure