GPU accelerated GMM supervectors for speaker and language recognition

Computing supervectors from many sliced utterance feature vectors as the inputs to support vector machine is used in many state-of-art systems for speaker and language recognition. This feature recombined method can achieve very well recognition results, but is also very time-consuming. By analyzing the supervectors computation procedure, we found great data-parallel potential. We can use vector/matrix linear algebra operations to compute supervectors. In this paper, we transferred the GMM supervectors computation from CPU to GPU and realized the speaker and language recognition systems based supervectors on the CPU-GPU hybrid platform. Compared to the implementation that used streaming SIMD extension (SSE) instructions on CPU, the supervectors computation on GPU was 63.8x faster. The CPU-GPU hybrid platform performed 67.4% and 44.5% accelerating ability respectively for speaker and language recognition systems. Using GPU to accelerate super-vectors computation, we can deal more a large number of utterance data in the same time.