Speech · Shipped · leaderboard-tier
Multilingual speech recognition
Dialect-aware recognition for languages where labelled audio is scarce and the incumbent systems are far larger than they need to be.
Arabic, across dialects
Our Arabic model is a roughly 115M-parameter acoustic model adapted in-house and fine-tuned on about 2,900 hours spanning Modern Standard Arabic and the Gulf, Egyptian, Levantine and Maghrebi groups.
Measured on all six sets of the public multi-dialect Arabic benchmark using its official code, it averages 25.08% word error rate. The systems around it are 2B–30B generative audio models that need a GPU; this one runs in real time on a CPU and ships as a single file.
Why size is the point
Comparable accuracy at roughly 18× fewer parameters decides whether a call centre can transcribe its own traffic on hardware it already owns, or whether transcription needs a GPU budget nobody approved.
A Tamil model follows the same approach, reaching 25.27% word error rate and 7.95% character error rate on the standard Tamil test split. More languages are in progress, chosen on the same basis each time: real speakers, scarce labelled audio, no efficient open model already serving it.
Models from this thread