---
pretty_name: GigaSpeech 3
language:
- ar
license: other
license_name: gigaspeech3-research-only
license_link: https://huggingface.co/datasets/speechcolab/gigaspeech3/blob/main/README.md#terms-of-use
task_categories:
- automatic-speech-recognition
annotations_creators:
- machine-generated
language_creators:
- found
multilinguality:
- monolingual
size_categories:
- 10M<n<100M
tags:
- audio
- speech
- arabic
- dialectal-arabic
- weak-supervision
- noisy-student-training
configs:
- config_name: EGY
  data_files:
  - split: train
    path: EGY/train/*.tar
- config_name: SAU
  data_files:
  - split: train
    path: SAU/train/*.tar
- config_name: SYR
  data_files:
  - split: train
    path: SYR/train/*.tar
- config_name: DZA
  data_files:
  - split: train
    path: DZA/train/*.tar
- config_name: IRQ
  data_files:
  - split: train
    path: IRQ/train/*.tar
---

# 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.

- **Dataset:** [GigaSpeech 3](https://huggingface.co/datasets/speechcolab/gigaspeech3)
- **Paper:** [arXiv:2611.02476](https://arxiv.org/abs/2611.02476)
- **Evaluation benchmark:** [GigaSpeechBench](https://huggingface.co/datasets/speechcolab/GigaSpeechBench)
- **Proofreading portal:** [VoxLabel](https://open-ai-eval.com)
- **Code:** [Antares-97/GigaSpeech3](https://github.com/Antares-97/GigaSpeech3)
- **Related datasets:** [GigaSpeech](https://huggingface.co/datasets/speechcolab/gigaspeech), [GigaSpeech 2](https://huggingface.co/datasets/speechcolab/gigaspeech2)

## Release version

This repository contains the complete WebDataset release: `<DIALECT>/train/NNN.tar`, with one WAV and one JSON per sample, covering all five dialects in 886 shards. The migration from the legacy WAV-only layout is complete. The original v1 WAV-only tag has been retired.

## 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](https://arxiv.org/abs/2611.02476) and rounded to whole hours; the total is reported separately in the paper.*

The WebDataset EGY subset contains the expanded release inventory: **1,912,374 segments / 11,166.6 hours**. The table above reports the earlier 8,533-hour EGY inventory used in the manuscript; the other four dialects are unchanged.

## Dataset Structure

### WebDataset layout

This repository serves self-contained WebDataset shards. Audio and metadata are packaged together:

```text
giga3/
├── README.md
├── DZA/
│   └── train/
│       ├── 001.tar
│       ├── 002.tar
│       └── ...
├── EGY/
│   └── train/
├── IRQ/
│   └── train/
├── SAU/
│   └── train/
└── SYR/
    └── train/
```

Each tar contains flat, same-key WAV/JSON pairs (no directory prefix):

```text
EGY/train/001.tar
├── <id>.wav
├── <id>.json
├── <id>.wav
├── <id>.json
└── ...
```

Each JSON metadata record contains:

| Field | Description |
|---|---|
| `id` | Sample identifier; also the tar member basename |
| `duration` | Duration in seconds |
| `text` | Weakly labeled transcript |
| `dialect` | Dialect code (`DZA`, `EGY`, `IRQ`, `SAU`, or `SYR`) |
| `shard` | Three-digit shard number, such as `"001"` |

Shard numbering is three digits and matches the tar filename: JSON `"shard": "001"` corresponds to `001.tar`. WAV files are 16 kHz mono PCM_16. There is no separate Parquet file in the WebDataset release; the JSON travels with each sample inside the tar.

The WebDataset layout is the only currently served format. Human-transcribed evaluation data are hosted separately in GigaSpeechBench.

## Example Usage

```bash
pip install -U datasets huggingface_hub
```

### Load the WebDataset version

```python
from datasets import Audio, load_dataset

REPO_ID = "speechcolab/gigaspeech3"
ds = load_dataset(
    REPO_ID,
    "SAU",
    split="train",
    streaming=True,
).cast_column("wav", Audio(decode=False))

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

You can also download one shard and pass its local path to `load_dataset`:

```python
from huggingface_hub import hf_hub_download

archive_path = hf_hub_download(
    repo_id="speechcolab/gigaspeech3",
    repo_type="dataset",
    filename="EGY/train/001.tar",
)
```

Individual shards can be several gigabytes. All WebDataset shards for a dialect are matched by `<DIALECT>/train/*.tar`.

The original WAV-only tag has been retired; this repository is now fully WebDataset.

## Proofreading

A proofreading portal is available at [open-ai-eval.com](https://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](https://huggingface.co/datasets/speechcolab/gigaspeech3/discussions), without publicly disclosing sensitive information.

## Citation

Please cite the paper when using this dataset:

```bibtex
@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}
}
```