Resumen de Ultra low-power, high-performance accelerator for speech recognition
Automatic Speech Recognition (ASR) is undoubtedly one of the most important and interesting applications in the cutting-edge era of Deep-learning deployment, especially in the mobile segment. Fast and accurate ASR comes at a high energy cost, requiring huge memory storage and computational power, which is not affordable for the tiny power budget of mobile devices.
Hardware acceleration can reduce power consumption of ASR systems as well as reducing its memory pressure, while delivering high-performance.
In this thesis, we present a customized accelerator for large-vocabulary, speaker-independent, continuous speech recognition. A state-of-the-art ASR system consists of two major components: acoustic-scoring using DNN and speech-graph decoding using Viterbi search. As the first step, we focus on the Viterbi search algorithm, that represents the main bottleneck in the ASR system.
The accelerator includes some innovative techniques to improve the memory subsystem, which is the main bottleneck for performance and power, such as a prefetching scheme and a novel bandwidth saving technique tailored to the needs of ASR.
Furthermore, as the speech graph is vast taking more than 1-Gigabyte memory space, we propose to change its representation by partitioning it into several sub-graphs and perform an on-the-fly composition during the Viterbi run-time. This approach together with some simple yet efficient compression techniques result in 31x memory footprint reduction, providing 155x real-time speedup and orders of magnitude power and energy saving compared to CPUs and GPUs.
In the next step, we propose a novel hardware-based ASR system that effectively integrates a DNN accelerator for the pruned/quantized models with the Viterbi accelerator. We show that, when either pruning or quantizing the DNN model used for acoustic scoring, ASR accuracy is maintained but the execution time of the ASR system is increased by 33%. Although pruning and quantization improves the efficiency of the DNN, they result in a huge increase of activity in the Viterbi search since the output scores of the pruned model are less reliable. In order to avoid the aforementioned increase in Viterbi search workload, our system loosely selects the N-best hypotheses at every time step, exploring only the N most likely paths. Our final solution manages to efficiently combine both DNN and Viterbi accelerators using all their optimizations, delivering 222x real-time ASR with a small power budget of 1.26 Watt, small memory footprint of 41 MB, and a peak memory bandwidth of 381 MB/s, being amenable for low-power mobile platforms.
