Gaussian Mixture Model Evaluation for Automatic Speech Recognition on GPU

Automatic speech recognition (ASR) is a very demanding computing task. Much research has been done into developing new techniques to improve the speed of ASR applications, including works leveraging hardware such as field programmable gate arrays (FPGAs) and graphics processing units (GPUs). In this thesis, a section of the ASR system, Gaussian mixture model (GMM) evaluation, was accelerated using GPU computing techniques. Profiling of software-based ASR programs revealed that GMM evaluation was one of the most time-consuming steps, indicating that the acceleration of this segment of the program could yield great benefits overall. Utilizing NVidia’s CUDA programming model, GPU code was developed in accordance with the methods of previous experiments. Once baseline code had been established, a variety of enhancements were attempted, including concurrent streams and dynamic parallelism. A speedup of approximately 19x over serial code was achieved for the implementation using concurrent streams. A technique using dynamic parallelism was also discovered which would improve memory utilization at slight cost to application speed. Many future applications of this code are possible, with several already being explored. To improve GPU performance even further, it could be beneficial to improve data organization in memory via the use of acoustic clustering algorithms. The GPU application could also be extended to include the graph search component of the ASR system by leveraging GPU breadth-first search techniques developed at the University of Illinois at Urbana-Champaign. By exploring these and other methods, ASR systems can be developed that will meet the demands of modern technology.Ope