---
language:
- th
license: cc-by-4.0
task_categories:
- automatic-speech-recognition
arxiv: 2601.13044
dataset_info:
  features:
  - name: audio_id
    dtype: string
  - name: sentence
    dtype: string
  splits:
  - name: test
    num_bytes: 162246
    num_examples: 1000
  download_size: 78444
  dataset_size: 162246
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test-*
---

# Gigaspeech2 Typhoon

[Project page](https://opentyphoon.ai/model/typhoon-asr-realtime) | [Paper](https://huggingface.co/papers/2601.13044) | [GitHub](https://github.com/scb-10x/typhoon-asr)

**Gigaspeech2 Typhoon** is a metadata-only reference dataset for Thai speech recognition benchmarking, specifically designed as an **Accuracy Track** for evaluating ASR models. The dataset contains 1,000 test samples with audio IDs and human transcriptions derived from the [Gigaspeech2](https://arxiv.org/abs/2406.11546) corpus. Each audio_id directly links to the original Gigaspeech2 dataset, allowing users to download the corresponding audio.

## Dataset Overview

- **Language:** Thai
- **Domain:** Audio from Gigaspeech2
- **Samples:** 1,000 test samples
- **Format:** Metadata-only (audio_id + sentence)
- **Task:** ASR Accuracy Evaluation (Accuracy Track)
- **License:** CC-BY 4.0 (inherited from [Gigaspeech2](https://arxiv.org/abs/2406.11546))
- **Original Source:** [speechcolab/gigaspeech2](https://huggingface.co/datasets/speechcolab/gigaspeech2)
- **Audio Access:** Download from original Gigaspeech2 using audio_id

## Dataset Structure

The dataset contains a single `test` split intended for evaluation purposes only.

### Data Fields

- **audio_id** (`string`): Unique identifier linking directly to [Gigaspeech2](https://arxiv.org/abs/2406.11546) audio samples (e.g., `63-63248-21`)
- **sentence** (`string`): Human-transcribed Thai text

> **Note:** This is a **metadata-only** dataset. Audio files are not included. To obtain the audio, download from [speechcolab/gigaspeech2](https://huggingface.co/datasets/speechcolab/gigaspeech2) using the audio_id.

### Data Statistics

- **Total samples:** 1,000
- **Average transcription length:** 48.16 characters
- **Min/Max transcription length:** 5 / 123 characters
- **Average word count:** 2.00 words
- **Min/Max word count:** 1 / 11 words

### Example

```python
from datasets import load_dataset

dataset = load_dataset("typhoon-ai/gigaspeech2_typhoon")

# Access the test split
test_data = dataset['test']

# View a sample
print(test_data[0])
# {
#   'audio_id': '63-63248-21',
#   'sentence': 'มันมาทางซ้ายแล้วก็พุ่งไปทางฝั่งขวามือข้างหน้ารถเขาอะฮะ'
# }

# To get the actual audio, download from original Gigaspeech2:
# original_gs2 = load_dataset("speechcolab/gigaspeech2", streaming=True)
# Then match using audio_id
```
## Attribution

This dataset is a **curated subset** derived from the original [speechcolab/gigaspeech2](https://huggingface.co/datasets/speechcolab/gigaspeech2) dataset. All audio samples in this dataset:

- Originate from the Gigaspeech2 Thai corpus
- Follow the Gigaspeech2 audio ID convention (e.g., `63-63248-21`)
- Maintain the same Creative Commons licensing as the source

### Original Dataset Citation

Please also cite the original Gigaspeech2 paper:

```bibtex
@article{gigaspeech2_2024,
  title={GigaSpeech 2: An Evolving, Large-Scale and Multi-domain ASR Corpus for Low-Resource Languages with Automated Crawling, Transcription and Refinement},
  author={Guan-Ting Lin and others},
  journal={arXiv preprint arXiv:2406.11546},
  year={2024}
}
```

**Verification:** All audio IDs in this dataset conform to the Gigaspeech2 naming convention and are traceable to the original corpus.

## Limitations

- **Test-only dataset:** Contains only 1,000 samples intended for evaluation, not for training
- **Metadata-only:** Audio files not included - users must download from original [Gigaspeech2](https://huggingface.co/datasets/speechcolab/gigaspeech2)
- **Clean audio focus:** Represents clean, academic-style speech from Gigaspeech2
- **Size:** Relatively small dataset suitable for benchmarking but not comprehensive coverage
- **Domain specificity:** Academic-style speech may differ from conversational or spontaneous speech

## Citation

If you use this dataset in your research or application, please cite our technical report:
[![arXiv](https://img.shields.io/badge/arXiv-2601.13044-b31b1b.svg)](https://arxiv.org/abs/2601.13044)

```bibtex
@misc{warit2026typhoonasr,
      title={Typhoon ASR Real-time: FastConformer-Transducer for Thai Automatic Speech Recognition}, 
      author={Warit Sirichotedumrong and Adisai Na-Thalang and Potsawee Manakul and Pittawat Taveekitworachai and Sittipong Sripaisarnmongkol and Kunat Pipatanakul},
      year={2026},
      eprint={2601.13044},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2601.13044}, 
}
```

## Contact

For questions or feedback about this dataset, please visit [opentyphoon.ai](http://opentyphoon.ai) or open an issue on the dataset repository.