Automatic Language Identification.
Final rept. Aug 72-Feb 74,
TEXAS INSTRUMENTS INC DALLAS
Pagination or Media Count:
Automatic language identification was based on the time-average log likelihood computed for certain preselected sound segments. These segments were isolated from thousands of candidate segments representimg sound transitions and steady-state sounds extracted from the training data. This selection procedure produced a set of segments useful in discriminating between pairs of languages, and was carried out for each of the ten language pairs for the five languages. Overall identification accuracies were estimated using 50 independent test speakers. Identification ranged from 60 to 100 for the language pairs. A decision strategy for identification among five languages used the pairwise identification results and yielded accuracies of 90, 85, 90, 30, and 0 for five languages. Overall accuracy was 64. Author