unfold-whistle-fr
A French fine-tune of Whistle, Cactus Compute's on-device speech model, for meetings. It is the model Unfold Note runs on the computer when French is the chosen spoken language. It is not a Whisper model.
- File:
whistle-fr.cact, 29.3 MB, 8 decoder layers, 4-bit weights and 8-bit activations (W4A8). It loads in the Cactus engine like the basewhistle.cact(tested withcactus-needle0.3.0, WASM). - Use it only for French. With language auto-detection it reads English as French (21 % WER on English synthetic
speech, against 5.4 % with English forced). Set the language to
fr.
Training
- Base:
Cactus-Compute/whistle, checkpointcheckpoints/whistle.safetensorsat revision741118687e1f. - Method: a LoRA adapter (rank 16, alpha 32) trained with
cactus whistle finetune(cactus-compute4.0.1, quantisation-aware: 4-bit weights during training), learning rate 1e-3, 3 epochs, batch 16. 90 % of the adapter is merged into the base (the update scaled by 0.9) before the 4-bit export. - Data:
- 22.75 hours of French business meeting recordings of the two Unfold Note founders, cut into 12,334 clips of at most 28 seconds, with the transcripts of a server speech recogniser as targets. Both founders consented to this use. The recordings are private and are not released.
- The 3,193 clips (10.3 hours) of the French training split of FLEURS (CC BY 4.0), mixed in so the model keeps general French: trained on meetings alone, it starts writing meeting talk into read speech.
Results
Word error rate (WER), lower is better, decoded with Unfold Note's own pipeline (pause cutting into pieces of at most 25 seconds, language forced to French).
| Test set | Base whistle.cact (b358ddad) |
unfold-whistle-fr |
|---|---|---|
| Held-out meetings: 22 meetings of one founder, 7.3 h, against the server transcripts | 39.1 % | 34.5 % |
| FLEURS French test, 676 clips | 33.5 % | 34.0 % |
| French synthetic speech, 205 s | 7.3 % | 5.9 % |
- On meetings, errors fall by 11.9 %; on the words that are names of people and organisations, from 58 % to 40 %.
- General French (FLEURS) is 0.4 point worse.
- Speed and memory are those of the base model on an Intel i7-14700F: real-time factor 0.17, about 160 MB per loaded engine (+15 to +40 MB against the 2-bit base).
- The meeting references are themselves speech recognition output: the meeting numbers measure agreement with it.
Limitations
- Trained on two people's meetings in one small company: names and terms frequent there are likelier outputs.
- Quiet audio, under about −42 dBFS, can be dropped by pause-based cutting before the model sees it; that is a property of the pipeline, not of the model.