DURATION-NORMALIZED FEATURE SELECTION FOR INDIAN SPOKEN LANGUAGE IDENTIFICATION IN UTTERANCE LENGTH MISMATCH

Spoken Indian language identification (SLID) plays a significant role in multilingual call center automation. The ability to identify the language of a very short-length utterance is crucial in a call center to enable route the call to an agent who can communicate in the native language of the caller. In this paper, we propose a duration normalized feature selection technique and show through extensive experimentation that this helps in improving the language identification, especially when the length of the spoken utterance is unknown a priori. We show that proposed duration normalized feature selection followed by output fusion of different classifiers perform best for utterance length mismatch condition. The relative improvement in accuracy from 31.5% to 99.0% and 10.4% to 25.9%, when trained with 30 s utterances and tested with 15 s and 0.2 s utterances, is achieved using a 150 duration normalized feature set.