Optimized Acoustic Likelihoods Computation for NVIDIA and ATI/AMD Graphics Processors

摘要: In this paper, we describe an optimized version of a Gaussian-mixture-based acoustic model likelihood evaluation algorithm for graphical processing units (GPUs). The these likelihoods is one the most computationally intensive parts automatic speech recognizers, but it can be parallelized and offloaded to GPU devices. Our approach offers significant speed-up over recently published approaches, because utilizes architecture in more effective manner. All recent implementations have been intended only NVIDIA graphics processors, programmed either CUDA or OpenCL programming frameworks. We present results both OpenCL. Further, developed implementation ATI/AMD GPUs. Results suggest that even very large models used real-time recognition engines on computers equipped with low-end laptops. addition, completely asynchronous management provides additional CPU resources decoder part LVCSR. enables us apply fusion techniques together evaluating many (10 more) speaker-specific models. technique parliamentary system where speaker changes frequently.