A Low-Resource ASR Back-End Based on Custom Arithmetic

Most contemporary ASR systems running on desktops use continuous-density HMMs (CHMM) with floating-point representations. It is important to reduce their memory and power requirements so that they can be more affordable for portable devices. In this paper, we propose a novel speech recognition back-end based on custom arithmetic, where all floating-point variables are represented by integer indices and all arithmetic operations are replaced by a sequence of table lookups. One critical issue associated with table lookups is what we call an accumulative variable whose dynamic range is either large or unpredictable. Such a variable would introduce much distortion if quantized to low precision, so that the table lookup would incur a great loss of information. We therefore explore different quantization structures dealing with this problem in likelihood evaluation, and present a normalization method for the Viterbi search to make the range of the forward probabilities predictable. Furthermore, we investigate several optimization algorithms on system-wide bit-width allocation. The best algorithm uses 80 Kbytes of tables to achieve all back-end operations with only a slight degradation in system performance. As a side effect, the offline storage for parameters is reduced by 80%, and the memory requirement for online computation is reduced by nearly 70%. Keywords—speech recognition, low resource, quantization, normalization, optimization