A Memory-Efficient Graph Structured Composite-State Network for Embedded Speech Recognition

There is a great demand to optimize the search engine of embedded speech recognition (ESR) system to make it applicable for low-resource portable devices. This paper focuses on the construction of a memory-efficient search space representation. To reduce the number of HMM models in prefix part, tree structured phonetic network is generally used. However, it still suffers from several kinds of redundancy, such as redundancy in suffix part, in state level and in topological structure. In order to eliminate such redundancy, we present a novel graph structured composite-state network. A comparison with traditional tree structured phonetic network shows that, our proposed network can reduce the memory footprint by a compression factor of 1.5 even with a relative speed up of 15.5% and without any loss in recognition accuracy.