Prosodic features in automatic language identification reflect language typology

Results from a prosody-based automatic language discrimination (LID) system suggest that the difficulties reported by other sites in incorporating prosodic information into LID systems may have been caused by their not using appropriate task-specific features. Running averages and correlations of prosodic features capturing syllable pitch and amplitude contours, duration and phrase location were evaluated by deriving a LLR function for each feature and language pair, then evaluating the effectiveness of that function as a discriminator. Data consists of speech in 11 languages (OGI database) representing a cross-section of traditional typological categories and relationships. Results show that prosody is highly useful in LID if complex perceptual events are broken down into simpler physical events and features are chosen based on task. Prosodic features can distinguish between language pairs as predicted by language typologies, suggesting that new languages can be classified using existing models of similar languages.