Abstract: This research investigates the automated identification of regional Swiss French accents using syllable-based prosodic features. By focusing on rhythm, intensity, and voice quality metrics like jitter and shimmer, the study demonstrates that a machine learning approach can achieve classification accuracies comparable to human performance, highlighting the subtle but distinct nature of prosodic variation in Swiss French.
The Challenge of Swiss French Accents
In linguistics, a sharp distinction exists between dialect (vocabulary) and accent (pronunciation). For Swiss French, regional variations are particularly subtle. Unlike languages where accents involve distinct vowel shifts, Swiss French accents are primarily defined by their "prosody"—the musicality, timing, and stress patterns of speech.
Because these differences are so nuanced, even humans struggle with identification; previous studies show human listeners only achieve about 43% accuracy in similar tasks. This paper aims to bridge this gap by creating an automated Regional Accent Identification (RAI) system, which is vital for developing natural-sounding voice assistants and improving speech recognition software for diverse speakers.
Methodology: The Syllable as a Unit
The researchers moved away from traditional phoneme-based analysis, focusing instead on the syllable as the basic unit of measurement. Their approach involved:
- Feature Extraction: Focusing on the three pillars of prosody: pitch, intensity, and rhythm.
- Voice Quality Metrics: Beyond standard pitch tracking, the authors included "Jitter" (micro-variations in frequency) and "Shimmer" (micro-variations in amplitude). While typically used in clinical settings, these metrics were hypothesized to capture unique regional "textures."
- DLOP Coefficients: Discrete Legendre Orthogonal Polynomials were used to mathematically model the shape of the pitch contour over a syllable.
- Classification: A Support Vector Machine (SVM) with a polynomial kernel was employed to categorize features into one of four regional classes.
Dataset Composition
The study utilized a subset of the PFC database (Phonologie du Français Contemporain), comprising 32 speakers (4 males and 4 females from each region). The regions included Geneva (GE), Martigny (MA), Neuchâtel (NE), and Nyon (NY). The audio consisted of a mix of journalistic reading and spontaneous monologues, totaling approximately 3 hours and 20 minutes of audio aligned at the syllable level.
Key Insights: Prosody and Identity
- Prosody is the Primary Signal: In Swiss French, the transitions between sounds are more telling than the sounds themselves. By focusing on the syllable, the model captures the rhythmic "heartbeat" of the regional accent.
- Voice Quality Matters: Jitter and Shimmer—features usually reserved for clinical voice analysis—provided significant clues for regional accent identification.
- Contextual Modeling: The researchers included features from preceding and following syllables, allowing the model to analyze how the "melody" of speech flows over time.
- Accumulative Information: Accent information is cumulative; the system performs significantly better on longer utterances, suggesting that regional identity is encoded in full sentence patterns rather than single words.
Experimental Results
The model using the expanded feature set (Set2, including jitter and shimmer) outperformed the basic set (Set1), reaching a total accuracy of 37.13%. This represents a 12.9% relative improvement over the baseline.
As shown in the results, the Nyon (NY) accent was the most distinct, with the model identifying it correctly 56.25% of the time. While a total accuracy of ~37% may seem low, it is remarkably close to the ~43% accuracy achieved by human listeners, highlighting the inherent difficulty of the task. Furthermore, for utterances with 20 or more syllables, the identification accuracy climbed to 45.72%.
Limitations and Future Work
The primary limitation is the sample size of 32 speakers, which risks over-fitting to individual voices. Future work will involve incorporating more speakers as the PFC project expands and combining this syllable-based prosodic approach with acoustic-based frameworks to create a hybrid system. The authors also intend to perform deeper phonetic analysis to isolate specific prosodic markers that contribute most to successful classification.
Summary updated