Prediction Network Architecture in RNN-T for ASR

The ChannelAugment technique improves robustness of multi-channel models

Dual-encoder architecture enables more accurate and natural speech interfaces

We present a framework for estimating acoustic properties of a signal using deep learning

Discover how speech recognizers can operate in conditions unseen in training data.

End-to-end learning can adapt the way computers perceive various aspects of speech.