Antares97/giga3

GigaSpeech 3

Scaling Multi-Dialect Arabic Speech Recognition with Weakly Labeled Data

GigaSpeech 3 is a large-scale Arabic speech recognition corpus containing approximately 10 million weakly transcribed speech segments and 39,000 hours of audio across five dialects: Algerian, Egyptian, Iraqi, Saudi, and Syrian Arabic. It supports research on dialect-specific and multi-dialect ASR, weak supervision, and low-resource speech recognition.

Format update — WebDataset. This repository uses self-contained WebDataset tar shards. The older audio/*.tar + metadata/*.parquet layout has been replaced by <DIALECT>/train/*.tar; each sample contains a WAV and its JSON metadata in the same tar.

• Dataset: GigaSpeech 3
• Paper: arXiv:2611.02476
• Evaluation benchmark: GigaSpeechBench
• Proofreading portal: VoxLabel
• Code: Antares-97/GigaSpeech3
• Related datasets: GigaSpeech, GigaSpeech 2

Dataset Description

Code	Dialect
DZA	Algerian Arabic
EGY	Egyptian Arabic
IRQ	Iraqi Arabic
SAU	Saudi Arabic
SYR	Syrian Arabic

Following the accompanying manuscript, dialect-specific labeling models are built using VietASR-style ASR-biased self-supervised learning and approximately 51–63 hours of manually transcribed speech per dialect. These models generate weak transcripts for dialect-matched web audio. Data selection and four rounds of noisy-student refinement for SAU and SYR produce the study's final training inventories.

The manuscript reports 23.86% macro-average WER for the final dialect-specific Zipformers on the five Arabic dialects of GigaSpeechBench, and evaluates transfer to MASC, Casablanca, and SADA. MSA speech used in the manuscript's multi-dialect experiments is separate from this five-dialect release.

Corpus Size

The five dialect subsets contain approximately 39,000 hours of weakly labeled Arabic speech.

Code	Dialect	Training hours
DZA	Algerian Arabic	11,461
EGY	Egyptian Arabic	8,533
IRQ	Iraqi Arabic	10,162
SAU	Saudi Arabic	4,512
SYR	Syrian Arabic	4,361
Total	Five dialects	39,028

Training durations are reported in the GigaSpeech 3 paper and rounded to whole hours; the total is reported separately in the paper.

Dataset Structure

Audio and metadata are packaged together as self-contained WebDataset shards:

GigaSpeech3/
├── README.md
├── DZA/
│   └── train/
│       ├── 001.tar
│       ├── 002.tar
│       └── ...
├── EGY/
│   └── train/
├── IRQ/
│   └── train/
├── SAU/
│   └── train/
└── SYR/
    └── train/

Each sample contains a WAV audio member and a JSON metadata member:

SYR/train/001.tar
├── sample_000001.wav
├── sample_000001.json
├── sample_000002.wav
├── sample_000002.json
└── ...

The JSON metadata contains:

Field	Description
id	Sample identifier
text	Weakly labeled transcript
duration	Duration in seconds
dialect	Dialect code
shard	Three-digit shard number, such as "001"

WAV files are original 16 kHz mono speech and are copied without transcoding. The JSON is the per-sample metadata record; there is no separate Parquet file in this WebDataset release.

Each dialect is exposed as a train configuration. Human-transcribed evaluation data are hosted separately in GigaSpeechBench.

Example Usage

pip install -U datasets huggingface_hub

Load a WebDataset shard directly:

from datasets import Audio, load_dataset

REPO_ID = "Antares97/giga3"
ds = load_dataset(
    "webdataset",
    data_files=f"hf://datasets/{REPO_ID}/SAU/train/*.tar",
    split="train",
    streaming=True,
).cast_column("wav", Audio(decode=False))

sample = next(iter(ds))
print(sample["json"])
print(sample["wav"]["bytes"][:4])  # RIFF

You can also download one tar with hf_hub_download and pass its local path to load_dataset("webdataset", data_files=..., split="train"). Individual shards can be several gigabytes.

Proofreading

A proofreading portal is available at open-ai-eval.com (VoxLabel); access may require an account.

Limitations

Transcripts are machine-generated weak labels and may contain recognition or normalization errors. Country-level dialect labels are coarse; recordings may contain MSA or code-switching. Web-source coverage and labeling models may introduce demographic, topic, and transcription biases. Use independently transcribed data for evaluation.

Terms of Use

Consistent with the manuscript, this corpus and its derived transcripts and metadata are provided for scientific research and evaluation only. This release does not grant commercial-use permission. Rights in underlying recordings remain with their respective rights holders, and applicable source licenses and platform terms continue to apply. Software recipes are licensed separately.

For questions or removal requests, open a dataset discussion, without publicly disclosing sensitive information.

Citation

Please cite the paper when using this dataset:

@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}
}