Language identification using noisy speech

This paper describes experiments in automatic identification of spoken languages using recordings of noisy radio signals as a data base. Prior efforts used uncorrupted speech; we selected techniques that we believed would be robust in noise. One technique attempted to distinguish languages by applying a classical quadratic classifier to prosodic features extracted from pitch and energy contours. Another was designed to exploit the frequency of occurrence of characteristic sounds using formant locations to represent the sounds, and using a vector-quantization distortion measure as the basis for language decisions. The techniques were required to make decisions based on speech segments of a few seconds duration. Our final tests were conducted on over 4 hours of previously unprocessed speech. Three languages, each from a different major language group, were used for development and testing. Allowing 11% false rejection (no decision), we achieved 64% correct identification with short speech segments. Our plans include the application of Markov modeling techniques to language identification.