Is there plan for uni-directional models?
Are pull request for uni-directional models welcome? What would be the requirements?
Motivation:
A left-to-right uni-directional models will allow ASR to be used in real-time applications
because they allow decoding speech without seeing the whole audio in advance so the latency is minimal. They decode as the user speaks.
Thank you
Oplatek
PS: Some related questions might be
- RTF: One should need RF < 1.0 for real time ASR Inference time run speeds
- Similar questions on architecture but not very helpful answers Any Good Architecture for continous Inference