Using End-to-end Multitask Model for Simultaneous Language Identification and Phoneme Recognition

Correlated speech tasks can learn related information from each other to enhance the performance. In this paper, a unified model based on end-to-end DNN and multi-task learning is proposed, which can simultaneously realize language identification and phoneme recognition. As a special feature of this model, we use the syllable, a larger unit in linguistics, to model temporal patterns and multiple cues to improve the robust and accuracy of language identification/phoneme recognition, including phonetic, prosodic and phonotactic. These experimental results on three speech corpuses shown that the unified model and multiple cues achieves the accuracy improvements. In addition, the experimental results on NIST LRE07 shown that our method outperforms the other methods in the task of language identification.