Doesn't really matter for the end result though

Yep exactly! Streaming models have to run faster, because they're more compute heavy. Some compute from every batch is wasted when the next set of frames come in. Streaming models also have higher WER than fully offline models.

Our models are more comparable to nemotron-3.5-asr-streaming in that regard and I think the medium/large sizes do perform quite favorably. For offline models, I don't think it's worthwhile for us to try and compete with parakeet v2/3. Those are already really good, and we're in the process of adding those to FUTO Voice.

These models are filling some gaps we see for the kinds of things we tend to work on. But broadly they can slot into a lot of places where there isn't accelerated compute (Raspberry pi 4, laptops, local client side transcription, etc).

The model cards list all of our data sources. They are all publicly available datasets that are commonly used in training most speech recognition models. It is good to take a wide variety of domains for robustness. Clean speech vs noisy backgrounds. Prepared speech vs spontaneous speech. Studio microphones vs cheap laptop microphones. More exposure is better.

For raw number of hours, I would set a goal of 40k hours. That number worked well for us, although if you are training bigger models you might still benefit from more.

Yes, because MIT does not prevent public and private mass-surveillance initiatives from using our models against people.