Part of Advances in Neural Information Processing Systems 4 (NIPS 1991)
Yeshwant Muthusamy, Ronald Cole
We have developed a four-language automatic language identification sys(cid:173) tem for high-quality speech. The system uses a neural network-based segmentation algorithm to segment speech into seven broad phonetic cat(cid:173) egories. Phonetic and prosodic features computed on these categories are then input to a second network that performs the language classification. The system was trained and tested on separate sets of speakers of Ameri(cid:173) can English, Japanese, Mandarin Chinese and Tamil. It currently performs with an accuracy of 89.5% on the utterances of the test set.