omniasr-ctc-300m-waxal-orm

Fine-tune of Meta's Omnilingual ASR (wav2vec2_asr (CTC)) on Oromo conversational speech from the WAXAL corpus, part of a 19-language WAXAL×Omnilingual ASR benchmark.

Results (held-out WAXAL test set)

Metric	Score
WER	25.8%
UER (unit/char error rate)	6.2%

Fine-tuned for 5,000 steps from facebook/omniASR-CTC-300M on 2×H200.

Usage

# pip install git+https://github.com/facebookresearch/omnilingual-asr.git
from pathlib import Path
import torch
from huggingface_hub import snapshot_download
from fairseq2.data.tokenizers.hub import load_tokenizer
from omnilingual_asr.models.inference.pipeline import ASRInferencePipeline
from fairseq2.models.wav2vec2.asr.hub import get_wav2vec2_asr_model_hub as get_hub

ckpt = snapshot_download("waxal-benchmarking/omniasr-ctc-300m-waxal-orm")
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
dtype = torch.bfloat16 if device.type == "cuda" else torch.float32

hub = get_hub()
config = hub.get_arch_config("300m")
model = hub.load_custom_model(Path(ckpt) / "model", config, device=device, dtype=dtype)
tokenizer = load_tokenizer("omniASR_tokenizer_v1")

pipe = ASRInferencePipeline(model_card=None, model=model, tokenizer=tokenizer, device=device, dtype=dtype)
texts = pipe.transcribe(["your_audio.flac"])  # CTC needs no language token
print(texts)

Audio should be mono 16 kHz (the pipeline resamples if needed); .transcribe() also accepts a list of raw bytes or np.int8 arrays. Keep clips under 40 s.

Checkpoint format

Native fairseq2 sharded checkpoint (model/pp_00/tp_00/sdp_00.pt + model.yaml) — not a transformers model, so AutoModel will not load it. Load with the omnilingual_asr / fairseq2 libraries, pointing the ASR recipe's model.path at the downloaded model/ directory with model.family=wav2vec2_asr, model.arch=300m, tokenizer=omniASR_tokenizer_v1.

Citation

This model accompanies the WAXAL ASR Benchmark (arXiv:2606.02375).

@article{waxalnet2026,
  title  = {The WAXAL ASR Benchmark: Fine-Tuned Edge Models Across 19 African Languages},
  author = {Olufemi, Victor Tolulope and Babatunde, Oreoluwa and Njema, Ramsey and
             Gbotemi, Bolarinwa and Yen, Wanchi Lucia and Uzodinma, John and
             Ajayi, Sunday and Williams, Oluwademilade and Moshood, Kausar and
             Anyaele, Innocent Elendu and Arefaine, Akebert Tesfahunegn and
             Hunzwi, Candace and Daniel, Wongel Dawit and Namuganga, Emmilly Immaculate and
             Kadima, Cleophas and Bahizire, Athanase Biluge and Ranaivoson, Onitsiky and
             Aaron, Emmanuel and Ladislaus, Nicholaus Dismas and Muhammed, Idris and
             Simenya, Jonathan Enoch and Koome, Martin and Endaylalu, Matewos Tegete and
             Adeyemo, Peter Ifeoluwa and Birindwa, Hondi Prisca and Eze-Mbey, Ukachi Agnes and
             Oduro-Yeboah, Yacoba and Aremu, Toluwani and Adjovi, Pericles and
             Ngueajio, Mikel K and Mitra, Prasenjit},
  year   = {2026},
  note   = {arXiv preprint arXiv:2606.02375}
}

Acknowledgements

We thank the native-speaker contributors for their language expertise and evaluation support. This work was supported by Lynguallabs (compute, researchers & storage), Open Token (compute resources), and CMU Africa (researchers & native speakers).