GTC-DC 2019: Accelerating Speech to Text Using Kaldi

GTC-DC 2019: Accelerating Speech to Text Using Kaldi

Levi Barnes, NVIDIA; Nigel Cannings, Intelligent Voice
NVIDIA and Intelligent Voice have been accelerating the Kaldi speech recognition framework using CUDA.This has been a continued effort and progress has been received with interest at previous GTC presentations.In these presentations amazing single GPU performance was demonstrated by showing LibriSpeech processing at 3500x real-time but multi-GPU scalability was limited.Recent developments include additional performance gains and the acceleration of feature extraction which greatly improved multi-GPU performance leading to near linear scalability on DGX-2. In addition to presenting these performance gains NVIDIA, Intelligent Voice, and HPE have partnered to bring speech recognition to the mass markets.For real world and multi-language applications, Intelligent Voice has more than 20 models that are now fully accelerated in CUDA.To better serve the speech market NVIDIA, Intelligent Voice, and HPE utilising their partnership to offer a pre-configured speech processing solution, details of which will be shared at this talk.While Kaldi is just a framework, IV provides an enterprise-strength, fully scalable speech platform to meet the most demanding applications in government, financial services and other commercial applications.