How can we use it for streaming ASR, say for models like nemo parakeet or openai whisper?

[programindz]

I would be really interested to see TorchStream being used in current SOTA models for ASR.

[CorentinJ]

That's a good idea, thanks for the suggestion.

I believe ASR models are often designed with streaming in mind because they are typically used in live applications:

• Parakeet is based on the fast conformer architecture which is natively streamable within Nemo, even if that demo shows a different model.
• Whisper is mainstream enough to have open source streaming implementations such as this one.

Nonetheless, it would be valuable insight to tackle either of these models with torchstream, to see how it handles a totally different use case with the challenge of an output format (i.e. the transcript with timestamps) that is not sliding-window based.

I'll give it a shot in the upcoming weeks, hopefully leading to a demo