Cross linguistic comparison of Mandarin and English EMA articulatory data

Sheng Li, Lan WangInterspeech 2012

TL;DR

This paper tackles the challenge of improving English pronunciation for Mandarin speakers by analyzing articulatory data. By comparing the precise movements of the tongue and lips during speech in both languages, the researchers identify specific phoneme pairs that are acoustically similar but articulately distinct, offering a new avenue for targeted pronunciation training.

What Problem Does This Paper Address?

Learning a new language often involves transferring phonetic habits from one's native tongue, leading to mispronunciations. For Mandarin speakers learning English, certain English phonemes are either absent in Mandarin or have acoustically similar Mandarin counterparts, making them difficult to distinguish and produce correctly. Traditional pronunciation training often relies on auditory feedback, which can be insufficient when the core issue lies in the physical production of sounds. This paper aims to bridge this gap by using detailed articulatory data to pinpoint these problematic phoneme pairs and provide a basis for more effective computer-assisted language learning (CALL) systems.

Methods

The core of this research lies in comparing the physical movements of the speech articulators (tongue, lips, jaw) between Mandarin and English speakers.

  • Electromagnetic Articulography (EMA): This technique uses sensors attached to the articulators to record their precise 3D movements at a high sampling rate.
  • Corpus Design: The study utilizes newly collected EMA data from a native Mandarin speaker and existing data from a native English speaker, covering a range of phonemes and speech samples.
  • Head Motion Normalization: To isolate articulator movements from head movements, a specific normalization technique using coils on the nose bridge and skull is applied.
  • Key Frame Extraction: For each phoneme, "key frames" are identified, representing the peak articulatory position characteristic of that sound, along with "static frames" representing the resting state.
  • Procrustes-Based Speaker Normalization: This is a crucial step to account for individual differences in vocal tract anatomy and biomechanics. It's a linear transformation that scales, rotates, and translates the articulatory data of one speaker to match another, preserving the relative shape of the articulators.
  • Contrastive Analysis: The researchers first identify English phonemes that are missing in Mandarin.
  • Cross-Linguistic Comparison:
    • Empirical Comparison: Visualizations of articulatory movements for confusable phoneme pairs are generated.
  • Quantified Comparison: Multi-Dimensional Scaling (MDS) is used to create a quantitative phonetic space where phoneme distances reflect their articulatory dissimilarity. Hierarchical Clustering Analysis (HCA) is also employed to group similar phonemes.

Datasets

  • English EMA Dataset: Collected from a native English speaker (female, age 22, Florida, USA).
  • Mandarin EMA Dataset: Newly collected from a native Mandarin speaker (female, age 40, Beijing, China).
  • Both datasets were recorded using the Carstens EMA AG500 at 200 frames per second.

Key Insights and Ideas

  • Articulatory Differences Underlie Acoustic Similarities: Acoustically similar but mispronounced phoneme pairs often have significant differences in the physical movements of the tongue and lips. This highlights the importance of looking beyond sound to understand pronunciation errors.
  • Speaker Normalization is Key: Individual anatomical variations are a major hurdle in cross-linguistic comparisons. Procrustes analysis effectively addresses this by creating a standardized articulatory space.
  • Visualizing Articulatory Space Reveals Distinctions: MDS and HCA provide powerful tools to visualize and quantify the differences in how phonemes are produced across languages, making abstract articulatory data more interpretable.
  • Targeted CALL System Potential: By identifying these specific articulatory confusions, CALL systems can be designed to provide targeted feedback and exercises, leading to more efficient pronunciation improvement.

Key Results

The paper presents qualitative and quantitative comparisons of phoneme pairs.

  • Consonants: Figures 2 and 3 visually demonstrate articulatory differences for several English consonants missing in Mandarin (e.g., /θ/, /ð/, /v/, /z/) and their Mandarin equivalents. For instance, the paper notes that while some consonants share articulation points, differences in lip opening and tongue position are evident.
  • Vowels: Similarly, Figure 3 shows distinct tongue positions and lip openings for English vowels (e.g., /ɪ/, /ʊ/, /ɛ/, /ɔ/, /æ/) compared to their Mandarin counterparts. The study suggests that Mandarin speakers may struggle to differentiate lax and tense vowels, and find equivalents for vowels like /ɛ/ and /æ/ in Mandarin.
  • MDS and HCA Visualizations: Figure 4 shows that both HCA and MDS analyses support the findings, indicating significant articulatory differences between English and Mandarin phonemes, even when they sound similar.

Limitations and Future Work

  • Limited Data: The study used data from only one native speaker for each language. The authors acknowledge that this limits the generalizability of the findings and suggest that future work should involve larger, multi-speaker datasets.
  • Future Work: The authors propose extending this methodology to compare different accents within the same language and to incorporate the identified cross-linguistic articulatory knowledge into future CALL systems for improved pronunciation training.

Summary updated

BibTeX
@inproceedings{li2012cross,
  author = {Sheng Li and Lan Wang},
  title = {{Cross linguistic comparison of Mandarin and English EMA articulatory data}},
  booktitle = {Interspeech},
  year = {2012},
}
Figure