The implementation of a practical high performance Mandarin and Sichuan Dialect continuous speech recognition system for parcels checking task

Abstract

This paper presents a high performance real-time Mandarin and Sichuan Dialect speaker-independent continuous speech recognition system utilized for a post parcels checking task. The vocabulary of the system consists of 4500 Chinese place names and 1021 number strings. For Mandarin speech the recognition accuracies are achieved 98.9% for top-1 and 99.7% for top-3. For Sichuan Dialect speech, the recognition accuracies are achieved 98.6% for top-1 and 99.9% for top-3. Besides, the rejection method based online garbage model and speaker adaptation method are employed and integrated into the system to improve its robustness. A mixed start-end point detection algorithm is used. The system can work stably under a high-noise background environment with high performance.