Algorithms for data-driven ASR parameter quantization

There is fast growing research on designing energy-efficient computational devices and applications running on them. As one of the most compelling applications for mobile devices, automatic speech recognition (ASR) requires new methods to allow it to use fewer computational and memory resources while still achieving a high level of accuracy. One way to achieve this is through parameter quantization. In this work, we compare a variety of novel sub-vector clustering procedures for ASR system parameter quantization. Specifically, we look at systematic data-driven sub-vector selection techniques, most of which are based on entropy minimization, and others on recognition accuracy maximization on a development set. We compare performance on two speech databases, phonebook, an isolated word speech recognition task, and timit, a phonetically diverse connected-word speech corpus. While the optimal entropy-minimizing or accuracy-driven quantization methods are intractable, several simple schemes including scalar quantization with separate codebooks per parameter and joint scalar quantization with normalization perform well in their attempt to approximate the optimal clustering.