Antares97/gigaspeech3_model

GigaSpeech 3 Models

Averaged Zipformer2 transducer checkpoints for five Arabic dialects and a multi-dialect model with MSA. These models were trained on the GigaSpeech 3 dataset. Each model uses avg=5 checkpoint averaging.

Each folder includes its SentencePiece tokenizer:

• SAU/, DZA/, EGY/, IRQ/, SYR/: unigram_500.model
• MULTI_DIALECT/: unigram_1500.model; supports DZA, EGY, IRQ, SYR, SAU, and MSA tokens with predicted LID.

The .pt files are PyTorch state_dict objects. Model architectures are Zipformer2 transducer models trained with icefall.

Links

• Dataset: speechcolab/gigaspeech3
• Paper: GigaSpeech 3 (arXiv:2611.02476)
• Evaluation benchmark: GigaSpeechBench
• Code: Antares-97/GigaSpeech3
• Proofreading portal: VoxLabel

Citation

@misc{liu2026gigaspeech3,
  title = {GigaSpeech 3: Scaling Multi-Dialect Arabic Speech Recognition with Weakly Labeled Data},
  author = {Liu, Junzhe and Tu, Yujie and Yang, Yifan and Wang, Peng and Wang, Haoran and Lin, Guodong and Liao, Leyan and Zhang, Chengquan and Gao, Zhifu and Li, Xiangang and Cheng, Qinyuan and Qiu, Xipeng and Zhang, Wei-Qiang and Chen, Guoguo and Yu, Kai and Chen, Xie},
  year = {2026},
  eprint = {2611.02476},
  archivePrefix = {arXiv},
  primaryClass = {cs.CL},
  url = {https://arxiv.org/abs/2611.02476}
}