Combining mixture weight pruning and quantization for small-footprint speech recognition

Abstract

Semi-continuous acoustic models, where the output distributions for all Hidden Markov Model states share a common codebook of Gaussian density functions, are a well-known and proven technique for reducing computation in automatic speech recognition. However, the size of the parameter files, and thus their memory footprint at runtime, can be very large. We demonstrate how non-linear quantization can be combined with a mixture weight distribution pruning technique to halve the size of the models with minimal performance overhead and no increase in error rate.

Low-power architectures for automatic speech recognition

“…In this framework, speech recognition is the process of selecting between a (potentially infinite) set of alternate transcriptions for a sentence. Equivalently, we can think of this as a classification of the acoustic observation, where each class contains all the realizations of a particular sentence [27].…”

“…maximum likelihood estimation is used, the variance of the resulting probability estimates will be quite large, and the model will also assign zero probability to many plausible word sequences [27].…”

“…However, a fanout at the roots, similar to that at the leaves, is undesirable since the former are active much more often. Therefore, cross-word triphones at the root nodes are modeled using the dynamic triphone mapping technique [27]. It multiplexes the states of a single root HMM between triphones resulting from different phonetic left contexts.…”

“…As shown in Figure 2.8, only the predecessor word with the highest path score is propagated to the point where the language model score is applied. However it is possible, and in fact quite likely, that another predecessor word would have been preferred had the language model score been available at word entry [27]. In order to solve this problem, a multiple pass search strategy is proposed which is explained in the following section.…”