Which speech model can I run?
Pick your hardware and requirements, and you'll get a ranked shortlist of open-source models. Rankings favour real evidence (our measurements, then sourced statements) and flag every gap as unknown instead of guessing. The URL updates, so you can share your result.
How ranking works
- Measured by us We ran it on this hardware class (exact machine on the page).
- Sourced statement The model authors or third-party docs make a specific statement about this hardware, quoted with a link.
- Fits stated memory The model's stated memory need fits this device. Nobody has published a run on it yet.
- CPU supported, no device data CPU inference is documented, but there's no data for this device.
- Estimate from parameter count Our rough fp16 weight-size estimate from the parameter count fits. Treat it as a guess.
- Unknown Not enough information. Shown last and never hidden, because unknown doesn't mean it won't work.
Within a level, models with fewer unknowns and fewer license conditions rank higher, then faster measured speed, then smaller size on edge devices. Sponsors can't buy placement here. Methodology
Sponsor
Reach developers picking a speech model to self-host. Advertise here →
Sponsors never affect rankings, picker results or benchmark numbers.