m-aliabbas1/asr-noise-robust-en

---
annotations_creators:
  - expert-generated
language:
  - en
license: other  # Source datasets may have varying licenses
multilinguality:
  - monolingual
pretty_name: ASR Noise-Robust English Dataset
tags:
  - audio
  - speech
  - automatic-speech-recognition
  - asr
  - noise-robust
  - whisper
  - wav2vec2
  - english
  - mixed-noise
  - dataset-aggregation
task_categories:
  - automatic-speech-recognition
task_ids:
  - automatic-speech-recognition
dataset_info:
  features:
    - name: audio
      dtype: audio
    - name: text
      dtype: string
---

# ASR Noise-Robust English Dataset

`m-aliabbas1/asr-noise-robust-en` is a curated multi-source dataset designed for building **noise-robust automatic speech recognition (ASR)** systems in **English**. It aggregates clean and noisy speech samples from a variety of well-established public datasets to simulate realistic transcription challenges.

---

## 🗂️ Dataset Composition

This dataset includes preprocessed data from the following sources:

- [`Myrtle/CAIMAN-ASR-BackgroundNoise`](https://huggingface.co/datasets/Myrtle/CAIMAN-ASR-BackgroundNoise): real background noise
- [`distil-whisper/librispeech_asr-noise`](https://huggingface.co/datasets/distil-whisper/librispeech_asr-noise): LibriSpeech with synthetic white and pub noise
- [`hhoangphuoc/buckeye`](https://huggingface.co/datasets/hhoangphuoc/buckeye): conversational American English
- [`hhoangphuoc/switchboard`](https://huggingface.co/datasets/hhoangphuoc/switchboard): telephone-based spontaneous dialogue
- [`edinburghcstr/ami`](https://huggingface.co/datasets/edinburghcstr/ami): meeting transcripts (validation split, IHM config)

---

## 📐 Structure

Each entry consists of:
- `audio`: audio waveform (decoded)
- `text`: transcript (lowercase, no punctuation)

The dataset contains:
- Varying audio conditions (clean, mixed, real-world noise)
- Audio lengths trimmed randomly to between **2 and 6 seconds**
- Empty transcripts for noise-only samples (e.g., from CAIMAN)

All audio features are normalized using:
```python
Audio(sampling_rate=None)

---

🧹 Preprocessing

• Transcripts were normalized:

  • Lowercased
  • All punctuation removed

• Audio was trimmed randomly between 2–6 seconds

• All splits from individual datasets were merged into a flat dataset

---

💡 Use Cases

• Training/fine-tuning noise-robust ASR models
• Evaluating model performance on realistic acoustic conditions
• Augmenting clean ASR datasets for robustness training

Compatible with:

• Whisper
• Wav2Vec2
• Conformer models

---

📈 Example

from datasets import load_dataset

ds = load_dataset("m-aliabbas1/asr-noise-robust-en")
print(ds[0]["audio"])
print(ds[0]["text"])

---

📃 License

This dataset aggregates samples from datasets under various licenses (e.g., CC BY 4.0). Each source retains its original license. Intended for research use only.

---

🙏 Acknowledgements

Thanks to the creators of:

• LibriSpeech
• AMI Corpus
• CAIMAN-ASR
• Buckeye Corpus
• Switchboard Corpus
• Hugging Face Datasets community

---

✨ Citation

@misc{asr-noise-robust-en,
  title = {ASR Noise-Robust English Dataset},
  author = {Ali Abbas},
  year = {2025},
  url = {https://huggingface.co/datasets/m-aliabbas1/asr-noise-robust-en}
}