Speech
Jackrabbit and Amami, open speech models for Romanian
Jackrabbit writes Romanian down. Amami reads it aloud. Both are small enough for a CPU.
- word error rate, FLEURS Romanian test, 883 clips
- 5.69%
- word error rate, FLEURS Romanian test, 883 clips
- parameters in Jackrabbit
- 116M
- parameters in Jackrabbit
- weights; commercial licences through Invergent
- CC-BY-NC-4.0
- weights; commercial licences through Invergent
Surogate Speech is our line of small, open speech models for agents: one model family per job, Jackrabbit to listen and Amami to speak, shipped language by language. Romanian is the first.
Jackrabbit
Jackrabbit 110M turns recordings into cased, punctuated Romanian. On the 883 clips of the FLEURS Romanian test it has a 5.69% word error rate, with CTC decoding and a 4-gram language model, scored with the Open ASR Leaderboard runner.
A streaming version shows words while you speak and writes the final sentence when you pause, over HTTP or WebSocket.
Amami
Amami reads Romanian aloud for voice agents. It puts the stress where native speakers put it, reads numbers, dates and amounts in words, and runs on a CPU.
Licence
The weights are under CC-BY-NC-4.0 for research and non-commercial use. Commercial licences come through Invergent.
Every clip on the Labs page was generated by the models, and the evaluation transcripts are public on Hugging Face.