A novel VQ-based speech recognition approach for mobile terminals

The ability to deploy speech recognition in mobile terminals, such as cellular telephones, has been limited to simple tasks due to restricted computational power and available memory. In this paper we present a new acoustic model based on vector quantization (VQ), which is better-suited for the limitations of the mobile environment. The proposed model is contrasted against finite state vector quantization (FSVQ), a commonly used technique is such environments. Performance enhancements include temporal restrictions to state generation, a flexible number of vectors per state, and updates of the traditional state transition function to a higher order. Experimental results show that the new model improves the recognition rate and reduces the computation load.