Introduction
Presented at INTERSPEECH 2012, this research addresses the "black box" nature of traditional dialect identification (DID) systems. While modern machine learning can categorize dialects with high accuracy, it often fails to explain the underlying linguistic features driving those decisions. This paper introduces an Acoustic Pronunciation Model (APM) designed to automatically extract and rank interpretable phonetic rules—such as [r]-dropping or vowel shifts—directly from raw audio data.
Methodology: The Acoustic Pronunciation Model
The core innovation is the APM, a framework that bridges the gap between high-performance acoustic modeling and linguistic interpretability.
Architecture
The APM is built upon a Hidden Markov Model (HMM) that maps "reference phones" (canonical pronunciations from a standard dictionary) to "surface phones" (the actual acoustic realizations). Unlike systems that rely on discrete phone labels, the APM utilizes continuous acoustic features to capture subtle variations. It explicitly models three types of phonetic transformations: insertions, deletions, and substitutions.
Rule Ranking
To identify the most characteristic features of a dialect, the researchers employed Log-Likelihood Ratios (LLR). By comparing the likelihood of a specific phonetic realization under a dialect-specific model (e.g., AAVE) versus a baseline model, the system can pinpoint and rank the most "informative" rules.
Experimental Setup and Datasets
The study utilized the StoryCorps corpus, focusing on natural conversations to capture authentic dialectal variations. The dataset comprised approximately 36.7 hours of audio, balanced for age and gender, categorized into two groups: African American Vernacular English (AAVE) and Standard American English (Non-AAVE).
Key Results and Insights
The APM demonstrated that explicit rule modeling provides a unique signal for dialect detection that standard systems often overlook.
Performance and Fusion
The APM achieved an Equal Error Rate (EER) of 10.73%, which is competitive with state-of-the-art GMM-SDC systems (10.86%). However, the most significant gains occurred during model fusion. Combining the APM with a phonotactic system (PPRLM) reduced the EER from 13.45% to 7.2%, suggesting that the APM captures complementary information.
Interpretable Rule Discovery
Beyond classification, the system successfully retrieved known linguistic rules and even refined existing knowledge. For instance, while linguists have long noted [ay] degliding (e.g., "my" sounding like "ma"), the APM discovered that this phenomenon is highly context-dependent, occurring more frequently when preceded by specific consonants like [w] or [m].
The system's robustness was further proven in unsupervised settings, where it performed nearly as well without word transcripts (45.02% F-measure) as it did with them (47.27%).
Conclusion and Future Directions
This "informative" approach to DID moves the field toward systems that can assist human experts rather than just providing a label. Future applications include forensic phonetics, where distinguishing between individual voiceprints and dialectal traits is crucial, and speech pathology, where the model could automatically characterize speech disorders by comparing patient realizations to a healthy reference model.
Summary updated