Quantizing mixture-weights in a tied-mixture HMM

We describe new techniques to significantly reduce computational, storage and memory access requirements of a tied mixture HMM based speech recognition system. Although continuous mixture HMMs offer improved recognition performance, we show that tied mixture HMMs may offer significant advantage in complexity reduction for low cost implementations. In particular, we consider two tasks: (a) connected digit recognition in car noise; and (b) subword modeling for command word recognition in a noisy office environment. We show that quantization of mixture weights can provide an almost three fold reduction in mixture weight storage requirements without any significant loss in recognition performance. Furthermore, we show that by combining mixture weight quantization with techniques such as VQ-Assist, the computational and memory access requirements can be reduced by almost 60-80% without any degradation in recognition performance