Smart City

(Panotti's Ears) Real time Voice Analytics Co-processor

AP007

Abdul Moeed (National University of Science & Technology )

Jul 31, 2021 1731 views

(Panotti's Ears) Real time Voice Analytics Co-processor

Speech recognition models have been used extensively on various platforms to provide ease of use, digital smart assistance, and hands-free control. At the cutting edge of this technology, the use of hidden Markov models is common. To improve the computational efficiency of hidden Markov model-based speech recognition systems, various techniques are used, amongst which the Viterbi beam search algorithm is one of the best. However, for large vocabulary speech recognition models with larger beam widths, the beam search algorithm’s sparse matrix operations create a highly constrictive bottleneck. Traditionally GPUs have been used to accelerate such models but with the algorithm's not so parallel nature, GPUs don’t provide an efficient solution and power constraints of IOT devices completely rule them out for Edge level.In this project, we research and formulate an FPGA based Co-processor (RTL level abstraction) to accelerate sparse matrix operation of the beam search algorithm so it can be used on edge devices to revolutionize how we interact with IOT edge level devices.

Project Proposal


1. High-level project introduction and performance expectation

2. Block Diagram

3. Expected sustainability results, projected resource savings

0 Comments



Please login to post a comment.