After clicking “Watch Now” you will be prompted to login or join.
Scaling Deep Learning for Automatic Speech Recognition
Jacob Kahn, Facebook | Vineel Pratap, Facebook | Vitaliy Liptchinsky, Facebook AI Research
GTC 2020
We'll discuss challenges of scaling automatic speech recognition (ASR) workloads with wav2letter++, a fast C++ toolkit for ASR. We'll introduce distributed training techniques used to achieve almost linear scalability and compare wav2letter to other popular ASR toolkits. Constant increase in model and dataset sizes, along with current trends toward unsupervised and semi-supervised learning, require squeezing out every bit of performance. In addition to distributed training, we'll cover other approaches for faster training and for training large models.