jsbeaudry/Oswald-ASR-0.6-V1-GGUF

Oswald-ASR-0.6-V1 — GGUF

Try it in your browser: The Trio Serie Space runs Oswald with the other two models of the series: the Oswald tab transcribes a recording or a file (the Space runs the full-precision model; these files are for llama.cpp). More on the Oswald page of thetrio.space.

Formerly jsbeaudry/qwen3-asr-0.6b-kreyol-v2-GGUF; the file names keep the old name so existing downloads keep working.

Haitian Creole speech recognition for llama.cpp. Converted from jsbeaudry/Oswald-ASR-0.6-V1.

The base model, Qwen/Qwen3-ASR-0.6B, does not support Haitian Creole — its card lists 30 languages and ht is not among them. Untrained it scores WER 90.1 on the test set below, writing Kreyòl as French. This is the fine-tune that taught it the language.

quant	decoder	projector	pair
Q8_0	767.5 MB	198.3 MB	965.8 MB
Q4_K_M	461.8 MB	198.3 MB	660.1 MB

Both decoders use the same Q8_0 projector. Q8_0 is the one to use unless size is binding.

Accuracy

Measured on 875 held-out clips with human-made transcripts, beam 4, with the PyTorch model:

	WER	CER
Oswald-ASR-0.6-V1 (this model)	9.8	3.6
v1	11.3	4.1
Qwen/Qwen3-ASR-0.6B base	90.1	41.3
oswald-large-v3-turbo-m3 (whisper large-v3-turbo)	9.5	3.9

Per domain, against v1 — every domain improved, none regressed:

domain	v1	v2
studio_legal	5.4	3.4
bible_read	5.6	4.0
studio_finance	5.4	4.1
studio_health	7.2	5.7
studio_numbers	7.8	6.4
studio_social	8.3	7.1
read_cmu	7.6	7.1
radio (machine transcripts)	13.3	12.0
studio_history	20.9	16.6
youtube	26.1	24.6

Two honest qualifications. The whisper comparison is against that model's own test set, not this one, so read it as "has caught up on CER" rather than a like-for-like row. And these numbers are the PyTorch model's — the GGUF was verified to transcribe correctly through llama-mtmd-cli before publishing, but its WER after quantisation has not been measured separately and may differ.

Speed on a laptop (M3 Pro, Q4_K_M, llama.cpp): a median 0.21 s per clip, 90% under 0.24 s, on 120 sentences spoken by Fatiman-TTS-v1 (WER 9.8 / CER 2.6 on those).

Use

llama-mtmd-cli -m qwen3-asr-0.6b-kreyol-v2-Q8_0.gguf \
  --mmproj mmproj-qwen3-asr-0.6b-kreyol-v2-Q8_0.gguf \
  --audio clip.wav -p '' --temp 0

The model writes its own language tag, so output begins language Haitian Creole<asr_text>.

Training

146.1 h across 16 datasets: a read Bible (capped at 35 h so one narrator cannot dominate), Radio Haïti-Inter broadcast speech, CMU LTI read speech, validated YouTube, studio prompt sets, and four synthetic TTS corpora at 18% of the mix. LoRA r=64 on the decoder, audio encoder frozen, 3 epochs with early stopping on validation loss.

On the synthetic data. Training on it alone as a continued fine-tune was measured on this same test set and made the model worse — 11.4 → 12.6 WER, and youtube degraded from 26.6 to 31.0. It learned numbers and forgot spontaneous speech. The same corpora mixed into one run at 18% improved every domain instead. Mix synthetic in; do not continue-train on it alone.

Converting this yourself

The transformers-native repo cannot be converted directly. llama.cpp routes this architecture through its Qwen3-Omni handler, which reads config["thinker_config"]["audio_config"] and expects thinker.* tensor names; the -hf layout has audio_config at the top level and model.* tensors. scripts/qwen3_asr/to_legacy.py rewrites it — four renames plus untying lm_head — and validates all 612 tensor names against Qwen/Qwen3-ASR-0.6B before writing, so a silent mismatch is not possible.