Memory Efficient and Fast Speech Recognition System for Low-Resource Mobile Devices

摘要

In this paper, we consider practical issues such as memory efficiency and fast decoding to make continuous density hidden Markov model (CDHMM)-based large vocabulary speech recognition system work on resource limited mobile devices. Particularly, we focus on memory efficient acoustic modeling and fast state likelihood computation. The proposed techniques are implemented in a speaker-independent Korean speech recognition system running on a Personal Digital Assistant (PDA) with a 32-bit fixed-point processor operating at 400MHz. The system uses 0.5MB memory for representing 28448 Gaussians and it runs at 2.54xRT without serious degradation of accuracy on 10k phonetically optimized words recognition task domain.