On using prosodic cues in automatic language identification

Presents an effort to explore the utility of prosodic information in language identification/discrimination (LID) tasks. We present our model and results from pair-wise LID tasks with English, Spanish, Japanese and Mandarin using multi-speaker elicited spontaneous speech and a selected set of prosodic parameters. These language represent four different types of languages, varying in pitch use and timing. Parameters were designed to capture pitch and amplitude contours on a syllable-by-syllable basis, and to be insensitive to overall amplitude, pitch and speaking rate. Results show that prosodic cues alone can distinguish between some language pairs with results comparable systems, indicating that prosodic parameters are highly useful in automatic LID. However the statistical relationships between a number of individual feature deduced from timing and pitch measurements are needed to begin to capture such complex perceptual events as rhythm. Strengths of individual prosodic paramaters and classes of parameters-primarily pitch, and secondarily duration and location-reflect differences between the four languages, mostly as expected based on the linguistic literature, suggesting that effective use of prosodic parameters is aided by a understanding of the relationships between physical measurements and perceived linguistic events