Efficient Embedded Speech Recognition for Very Large Vocabulary Mandarin Car-Navigation Systems

Automatic speech recognition (ASR) for a very large vocabulary of isolated words is a difficult task on a resource-limited embedded device. This pal)er presents a novel fast decoding algorithm for a mandarin speech recognition system which can simultaneously process hundreds of thousands of items and maintain high recognition accuracy. The proposed algorithm constructs a semi-tree search network based on mandarin pronunciation rules, to avoid duplicate syllable matching and save redundant memory. Based on a two-stage fixed-width beam-search baseline system, the algorithm employs a variable beam-width pruning strategy and a frame-synchronous word-level priming strategy to significantly reduce recognition time. This algorithm is aimed at an in-car navigation system in China and simulated on a standard PC workstation. The experimental results show, that the proposed method reduces recognition time by nearly 6-fold and memory size nearly 2-fold compared to the baseline system, and causes less than 1% accuracy degradation for a 200, 000 word recognition task.