kadirnar/arabic-stt-runs

arabic-stt research runs

Training runs of small (about 5-30M parameters) Arabic and Arabic + English speech recognition models, trained from scratch on a small amount of labeled audio. Every technique is tried as a one-change ablation against a frozen current best. These are research artifacts, not a product.

Research use only. Non-commercial. CC BY-NC-SA 4.0. The training data include SADA (CC BY-NC-SA 4.0, whose ShareAlike term applies to derived releases) and MGB-2, MGB-3, MGB-5 and NADI-2026, which are distributed for non-commercial research only (or with unstated terms, treated as non-commercial research). The weights are therefore released only for non-commercial research, under CC BY-NC-SA 4.0. Anyone using them must also respect the terms of every corpus listed below.

What is here

One directory per finished run, exp/<experiment>/<arm>/<run id>/:

file	contents
config.yaml	the resolved training config (model, data mix, optimizer, schedule, seed)
run.json, done.json	run identity: config hash, git commit of the training code, start / end, steps, training hours
model.json, data.json, env.txt	parameter counts and FLOPs, training-data filter counts and mix, software and GPU versions
averaged/model.pt, averaged/average.json	the final averaged fp32 weights ({"model": state_dict, "window": ...}) and the averaging window
cmvn.pt	the feature normalization statistics the run trained with
results/<decoder>.json	aggregate metrics only: WER / CER per evaluation set and subset, both leaderboard scorer versions
metrics.jsonl, train.log, stdout.log	training and validation curves, training logs

Not published: intermediate checkpoints, per-utterance outputs (references and hypotheses), any transcript, the training manifests and the decoding language models. The training code is not part of this repository; each run records the commit it was trained with (git_sha). The tokenizer named in config.yaml (assets/tokenizers/) is needed to decode the model's outputs.

Training data and licences

Each run's config.yaml (data.train) names the tiers it trained on; together the runs use these corpora:

corpus	source (Hugging Face)	licence
Common Voice 18 Arabic	MohamedRashad/common-voice-18-arabic	CC0-1.0
FLEURS (ar_eg, en_us)	google/fleurs	CC-BY-4.0
MASC	pain/MASC	CC-BY-4.0
SADA	MohamedRashad/SADA22 (+ CSVs of khaledalganem/sada2022)	CC-BY-NC-SA-4.0
MGB-2	MohamedRashad/mgb2-arabic	MGB-2 research licence (non-commercial; the official QCRI/mgb2 terms)
MGB-3	ArabicSpeech/MGB-3	research, unstated (treated as non-commercial)
MGB-5	ArabicSpeech/MGB-5	research, unstated (treated as non-commercial)
NADI-2026	UBC-NLP/NADI2026_subtask1.1_Robust_ASR	unstated (treated as non-commercial research)
LibriSpeech	openslr/librispeech_asr	CC-BY-4.0
AMI (IHM)	edinburghcstr/ami	CC-BY-4.0
VoxPopuli (en)	facebook/voxpopuli	CC0-1.0
People's Speech (clean)	MLCommons/peoples_speech	CC-BY 2.0-4.0 per source
Common Voice 17 English	fixie-ai/common_voice_17_0	CC0-1.0
ArzEn (Hamed et al. 2020)	ahmedsamirtarjama/MBZUAI_ArzEn_wav	unverified (the copy has none): research only
Mixat	sqrk/mixat-tri	CC-BY-NC-SA-4.0

Wave augmentation mixes, on the fly, MUSAN (OpenSLR 17, CC BY 4.0; Snyder, Chen and Povey 2015, arXiv:1510.08484) and the simulated room impulse responses of RIRS_NOISES (OpenSLR 28, Apache 2.0; Ko et al., ICASSP 2017). Neither is redistributed.

Evaluation-only data

Never trained on; their numbers appear in results/*.json as aggregates only:

• Casablanca (UBC-NLP/Casablanca, CC-BY-NC-ND-4.0) and the six Arabic test sets of the Open Universal Arabic ASR Leaderboard (test_lb: Common Voice 18, MASC clean / noisy, SADA, MGB-2, Casablanca).
• English: GigaSpeech dev / test (speechcolab/gigaspeech, non-commercial research terms), the Open ASR Leaderboard short-form sets, the dev / test splits of the English corpora above.
• Code-switching: the ArzEn dev split, Mixat test, ESCWA.CS (CC-BY-NC-4.0) and Perle (MIT).

The ctc_beam_lm results decode with a token-level KenLM built from FineWeb-2 web text (HuggingFaceFW/fineweb-2) and the training transcripts; that LM is not published.