Downloads · 30 days
0
DariusTheGeek/waxal-lin-omniasr-ctc-1b
waxal-lin-omniasr-ctc-1b is a automatic speech recognition model from DariusTheGeek. Use it when you need speech turned into text. It is set up for fairseq2. The card lists the license as apache-2.0.
OmniASR CTC-1B v2, fine-tuned for Lingala as part of the WAXAL ASR solution.
Downloads · 30 days
0
Access
Public
Updated Aug 19, 2026
Repo size
3.9 GB
Likes
0
Public
Click a slice to open those files.
.pt3.9 GB · 100%
From the Hugging Face model README
OmniASR CTC-1B v2, fine-tuned for Lingala as part of the WAXAL ASR solution.
Released artifact: top-3 uniform average of steps 2505 / 3006 / 4008. Decoded with greedy.
| File | Bytes |
|---|---|
model.pt | 3,902,903,831 |
omniASR_tokenizer_written_v2.model | 91,481 |
card.yaml | — |
requirements.txt | — |
Total weights: 3,902,903,831 bytes (3.90 GB).
Member of the Lingala TTIA fusion; the conservative-ROVER candidate over all four members is scored alongside the members themselves.
This model is one component of an ensemble solution and is not intended to be used alone. The routing, decoding, fusion and post-processing pipeline are in the solution repository.
This is a fairseq2 / omnilingual-asr checkpoint, loaded through the asset
card shipped alongside it (card.yaml). The solution repository wraps the whole
contract -- pinned environment, batch-size-1 decode, tokenizer wiring -- in one
CLI:
git clone https://github.com/DariusTheGeek/waxal-asr-solution
cd waxal-asr-solution && bash install.sh # pinned environments, ~15 min
python models/download_models.py --repo waxal-lin-omniasr-ctc-1b
.venvs/omni/bin/python inference/decode/omniasr.py \
--config configs/lin/ctc1b.yaml \
--audio path/to/wav_dir --output transcripts.csv
--weights accepts any directory holding this repo's files, e.g. the path
returned by huggingface_hub.snapshot_download("DariusTheGeek/waxal-lin-omniasr-ctc-1b").
requirements.txt in this repo pins the runtime alone; the environment lock
the release was verified under is env/requirements-omni.txt in the solution
repository.
| Parent | facebook/omniASR-CTC-1B |
| Language | Lingala |
| Fine-tuning data | Waxal Lingala/Shona supervised split (google/WaxalNLP) |
| Seed | 42 |
See https://github.com/DariusTheGeek/waxal-asr-solution for the environment locks, decode configuration and the exact
command that reproduces the submission end to end from audio.
apache-2.0, inherited from the parent model.