Introduction

finanfa-fon-asr-v1 is live and downloadable: finanfa-tech/finanfa-fon-asr-v1 (created 2026-11-04 ~09:38 UTC). Finanfa Tech fine-tuned Whisper Small (242M) on Fon / Fongbe (fon) speech from Benin, then exported GGML quantizations for whisper.cpp so you can transcribe offline on CPU.

This cycle matters because the Finanfa Fon stack sat on our soft-hold list as 401 / WIP. The ASR repo is now HTTP 200, ungated, and Apache-2.0 — still labeled version 1, de travail, but runnable today.

What Shipped

  • Vendor / product: finanfa-fon-asr-v1 — Finanfa Tech
  • Modality: open-weight automatic speech recognition (Fon / Fongbe, with French language token routing)
  • Open vs closed: open weights, Apache-2.0
  • Runnable today: compile whisper-cli, download a .bin, force -l fr, transcribe 16 kHz mono WAV
  • Concrete fact: two packs ship — finanfa-fon-asr-v1-q4_0.bin ~145 MB (faster / smaller) and finanfa-fon-asr-v1-q5_1.bin ~190 MB (slightly more faithful); independent bible-read bench (124 clips, 10.7 min) reports Q5_1 WER 51.2% with tones ignored vs 82.2% with tones kept, at 0.24× realtime on a MacBook M3 CPU (4 threads)

The card’s own 20-clip development WER (21.99%) is optimistic and tiny — Finanfa says so. Treat the independent tone-sensitive numbers as the honest floor for careful orthography.

Run It

ffmpeg -i audio.mp3 -ar 16000 -ac 1 -c:a pcm_s16le audio.wav
./whisper-cli -m finanfa-fon-asr-v1-q5_1.bin -f audio.wav -l fr -nt

-l fr is mandatory. The fine-tune routes Fon through Whisper’s French language token; without it, output is unreliable. -nt drops timestamps; -t 4 sets threads. Long audio should be split — Whisper still windows at ~30 s.

Sibling finanfa-fon-tts-v1 (Piper/VITS ONNX, ~63.5 MB, tone-aware) is also public now, but stays on soft-hold: non-commercial research-only + explicit WIP tone/symbol caveats. Use the ASR card if you need an Apache path today.

Notes for makers

  • Expect missing or incomplete tone marks, occasional ε vs ɛ, and model-added punctuation — plan a normalization step before scoring
  • Best fit right now: draft transcripts and keyword search, not tone-perfect archival text
  • Keep Finanfa TTS on the soft-hold list until the license/WIP story clears for commercial makers
  • No Gen CTA: Fon ASR does not map to the Flux / Grok / MiniMax-H3 / Seedream / Seedance Gen loop

Conclusion

If you need a CPU-first Fon speech-to-text pack you can actually download, Finanfa’s whisper.cpp GGML drop is the soft-hold unlock this cycle — start with Q5_1, force French language routing, and keep expectations calibrated to the independent WER.

—Anabel ♡