A Novel Large Vocabulary Continuous Speech Recognition Algorithm Combined with Language Recognition
一种联合语种识别的新型大词汇量连续语音识别算法

In this paper, a novel large vocabulary continuous speech recognition (LVCSR) algorithm combined with language recognition is proposed, and a real-time processing system is developed. This algorithm can make full use of phonetic hypotheses collected during decoding, and identify language types simultaneously. In a multilingual environment, this algorithm can not only take the place of a standalone language recognizer at a lower system overall computational cost, but also effectively cope with the case where target and non-target languages mix in a single utterance. It can significantly reduce speech recognition error introduced by non-target language, and avoid error accumulation which may mislead the subsequent decoding procedure. In order to tightly combine the content and language recognition into a unified decoding procedure, three different phone lattice reconstruction algorithms are also proposed to eliminate pronunciation and grammar restrictions introduced by the target language's dictionary and language model of the LVCSR decoder, and to encode lattices with richer phonetic information. Experiments show that the lattice reconstruction algorithms can significantly improve language recognition accuracy in the combined recognition. Evaluated on a Mandarin/English mixed conversational telephone speech corpus where Mandarin is the target language, the proposed algorithms reduced the recognition error introduced by non-target language by 91.76%, and achieved a character error rate of 54.98%.