Modelpublic

waxal-benchmarking/omniasr-llm-300m-waxal-all19

sourceHugging Facecc-by-nc-4.0updated 15d agoView on Hugging Face

1likes

Model Card

omniasr-llm-300m-waxal-all19

A all 19 WAXAL languages (multilingual) multilingual fine-tune of Meta's Omnilingual ASR (wav2vec2_llama, 300M-parameter shared wav2vec2 encoder + autoregressive Llama-style decoder), trained jointly on 19 WAXAL languages. Part of the WAXAL ASR benchmark's training-granularity study (monolingual vs. language-family vs. all-19 pooling).

—Base model: `facebook/omniASR-LLM-300M`
—Languages (19): Acholi, Akan, Amharic, Dagbani, Dagaare, Ewe, Fula, Ikposo, Lingala, Luganda, Masaaba, Malagasy, Nyankole, Oromo, Sidama, Shona, Soga, Tigrinya, Wolaytta
—Macro-averaged WER (this model): 32.9%  |  monolingual baselines: 31.8%

Training data

Fine-tuned on the pooled train splits of the WAXAL corpus for these languages (16 kHz mono; transcripts NFC-normalized and lower-cased, punctuation removed, phonemic diacritics/tone marks preserved). Total: 362,117 clips / 1857.7 hours.

Language	Train clips	Train hours
Acholi	4,108	25.8
Akan	10,107	55.0
Amharic	38,022	189.7
Dagbani	14,231	77.0
Dagaare	15,071	83.4
Ewe	15,052	79.5
Fula	19,132	99.7
Ikposo	14,415	82.1
Lingala	14,399	71.9
Luganda	5,455	37.3
Masaaba	6,865	39.3
Malagasy	18,526	94.7
Nyankole	6,783	40.8
Oromo	38,182	189.9
Sidama	38,752	192.7
Shona	14,109	79.7
Soga	6,219	40.0
Tigrinya	40,856	181.9
Wolaytta	41,833	197.3

Training procedure

Fine-tuned with the Omnilingual-ASR wav2vec2_llama recipe (fairseq2) on 2× NVIDIA H200 (DistributedDataParallel). All granularity conditions use an identical budget so the only variable is the language mixture.

Hyperparameter	Value
Base checkpoint	`facebook/omniASR-LLM-300M`
Tokenizer	`omniASR_tokenizer_v1` (SentencePiece, 9,812 units)
Training steps	5,000
Optimizer	AdamW
Learning rate	5e-5
Weight decay	0.01
Batch size	3,000,000 audio tokens/batch (dynamic)
Gradient accumulation	4
Precision	bfloat16 mixed
Audio length filter	0.5 s – 30 s
Hardware	2× H200 (DDP)

Evaluation

Scored on each language's held-out test split (utterances ≥ 1.5 s, matching the benchmark's filtered-test protocol). WER and CER computed with jiwer on NFC-normalized, lower-cased text (diacritics preserved). The Monolingual WER column is the corresponding per-language model (`omniasr-llm-300m-waxal-<iso>`) evaluated identically, for a same-protocol comparison.

Language	Code	WER	CER	Monolingual WER	Δ vs mono
Acholi	`ach_Latn`	33.2	13.7	33.3	-0.1
Akan	`aka_Latn`	29.1	9.8	29.2	-0.1
Amharic	`amh_Ethi`	26.9	10.7	25.0	+1.9
Dagbani	`dag_Latn`	34.0	11.6	31.2	+2.8
Dagaare	`dga_Latn`	32.2	13.0	29.0	+3.2
Ewe	`ewe_Latn`	29.8	9.4	29.0	+0.8
Fula	`ful_Latn`	36.5	9.5	39.1	-2.6
Ikposo	`kpo_Latn`	72.5	24.6	69.2	+3.2
Lingala	`lin_Latn`	29.9	12.5	30.4	-0.5
Luganda	`lug_Latn`	12.1	2.8	10.9	+1.1
Masaaba	`myx_Latn`	45.1	10.2	45.5	-0.5
Malagasy	`mlg_Latn`	12.9	3.4	11.4	+1.5
Nyankole	`nyn_Latn`	32.2	8.4	32.3	-0.1
Oromo	`orm_Latn`	22.8	6.0	20.9	+2.0
Sidama	`sid_Latn`	30.5	7.5	28.7	+1.8
Shona	`sna_Latn`	21.6	3.9	21.6	+0.0
Soga	`xog_Latn`	42.7	8.5	42.8	-0.1
Tigrinya	`tir_Ethi`	40.6	17.4	37.0	+3.7
Wolaytta	`wal_Latn`	40.0	11.7	36.7	+3.2

Usage

python

# 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 omnilingual_asr.models.wav2vec2_llama.hub import get_wav2vec2_llama_model_hub

ckpt = snapshot_download("waxal-benchmarking/omniasr-llm-300m-waxal-all19")
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
dtype = torch.bfloat16 if device.type == "cuda" else torch.float32

hub = get_wav2vec2_llama_model_hub()
model = hub.load_custom_model(Path(ckpt) / "model", hub.get_arch_config("300m"), device=device, dtype=dtype)
tokenizer = load_tokenizer("omniASR_tokenizer_v1")

pipe = ASRInferencePipeline(model_card=None, model=model, tokenizer=tokenizer, device=device, dtype=dtype)
# pass the target language's Omnilingual token, e.g. Acholi -> "ach_Latn"
texts = pipe.transcribe(["your_audio.flac"], lang=["ach_Latn"])
print(texts)

Language tokens for this model: Acholi ach_Latn, Akan aka_Latn, Amharic amh_Ethi, Dagbani dag_Latn, Dagaare dga_Latn, Ewe ewe_Latn, Fula ful_Latn, Ikposo kpo_Latn, Lingala lin_Latn, Luganda lug_Latn, Masaaba myx_Latn, Malagasy mlg_Latn, Nyankole nyn_Latn, Oromo orm_Latn, Sidama sid_Latn, Shona sna_Latn, Soga xog_Latn, Tigrinya tir_Ethi, Wolaytta wal_Latn. Audio should be mono 16 kHz (the pipeline resamples if needed); 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 omnilingual_asr / fairseq2 as shown above.

Citation

Part of the WAXAL ASR Benchmark (arXiv:2606.02375).

bibtex

@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 others},
  year   = {2026},
  note   = {arXiv preprint arXiv:2606.02375}
}

Acknowledgements

Supported by [Lynguallabs](https://lynguallabs.org/) (compute, researchers & storage), [Open Token](https://opentoken.global/) (compute), and [CMU Africa](https://www.africa.engineering.cmu.edu/) (researchers & native speakers).