Language identification using multi-core processors

Graphics processing units (GPUs) provide substantial processing power for little cost. We explore the application of GPUs to speech pattern processing, using language identification (LID) to demonstrate their benefits. Realization of the full potential of GPUs requires both effective coding of predetermined algorithms, and, if there is a choice, selection of the algorithm or technique for a specific function that is most able to exploit the GPU. We demonstrate these principles using the NIST LRE 2003 standard LID task, a batch processing task which involves the analysis of over 600 h of speech. We focus on two parts of the system, namely the acoustic classifier, which is based on a 2048 component Gaussian Mixture Model (GMM), and acoustic feature extraction. In the case of the latter we compare a conventional FFT-based analysis with IIR and FIR filter banks, both in terms of their ability to exploit the GPU architecture and LID performance. With no increase in error rate our GPU based system, with an FIR-based front-end, completes the NIST LRE 2003 task in 16 h, compared with 180 h for the conventional FFT-based system on a standard CPU (a speed up factor of more than 11). This includes a 61% decrease in front-end processing time. In the GPU implementation, front-end processing accounts for 8% and 10% of the total computing times during training and recognition, respectively. Hence the reduction in front-end processing achieved in the GPU implementation is significant.

► Application of graphics processing units (GPUs) to real problems in automatic language identification (LID).
► Exploitation of GPU requires effective coding of predetermined algorithms, and selection of algorithms that are most able to exploit the GPU architecture.
► Gaussian Mixture Model (GMM) probability calculations accelerated by factors of up to 64.
► Acoustic feature extraction using an FIR filterbank on a GPU runs up to 14 times faster than a conventional DFT-based analysis on a CPU, with no increase in LID error rate.
► Complete LID system with FIR-based front-end running on a GPU completes the NIST 2003 LRE evaluation in 16 h, compared with 180 h for a conventional DFT-based analysis on a CPU, with no increase in LID error rate.