Firoj112/nepali-asr-whisper

Nepali ASR Dataset (FLAC, Prepared for Whisper Fine-Tuning)

This dataset is a preprocessed and ready-to-use version of the OpenSLR Nepali Automatic Speech Recognition (ASR) corpus, repackaged and standardized to facilitate Whisper model fine-tuning and other speech-to-text experiments.

It provides audio-text pairs in Nepali language, with all audio stored as .flac files and transcriptions in Devanagari script. The dataset follows the Hugging Face datasets format for seamless use with Transformers, SpeechBrain, ESPnet, or fairseq pipelines.

---

🗂️ Dataset Summary

Split	Examples	Approx. Size
Train	2,273,832	130.9 GB
Validation	126,324	7.31 GB
Test	126,324	7.24 GB
Total	~2.52M	~145.48 GB

All audio files have been resampled or confirmed to use a 16 kHz sampling rate to align with Whisper’s input requirements.

---

🧾 Data Fields

Column	Type	Description
utt_id	string	Unique utterance ID
speaker_id	string	Speaker identifier
sentence	string	Transcribed Nepali sentence
audio	Audio (16 kHz)	FLAC audio waveform
__index_level_0__	int64	Internal dataset index

---

📦 Dataset Structure

After extraction, the dataset directory was organized as:

extracted/
 ├── asr_nepali_0/
 │   └── asr_nepali/data/
 │       ├── 00/
 │       │    ├── *.flac
 │       ├── 01/
 │       ├── ...
 │       ├── 0f/
 ├── asr_nepali_1/
 │   └── asr_nepali/data/
 │       ├── 00/ ... 0f/
 ...

A master metadata file (master.tsv) was programmatically generated to map all audio files and their corresponding transcripts, later split into train/validation/test subsets.

---

🧠 Intended Use

• Fine-tuning Whisper, Wav2Vec2, or Conformer models for Nepali ASR tasks.
• Benchmarking speech recognition and language modeling in Nepali.
• Supporting RAG (Retrieval-Augmented Generation) speech-text pipelines or multilingual model adaptation.

---

⚙️ Example Usage

from datasets import load_dataset

dataset = load_dataset("Firoj112/nepali-asr-flac", split="train")
print(dataset[0])
# {
#   'utt_id': 'utt_001',
#   'speaker_id': 'spk_12',
#   'sentence': 'नेपाल सुन्दर देश हो ।',
#   'audio': {'array': [...], 'sampling_rate': 16000}
# }

---

🧩 Technical Notes

• All .flac files were loaded using soundfile (sf.read) to ensure consistent audio arrays.
• Dataset split ratio: Train 90% / Validation 5% / Test 5%
• Audio and transcription alignment verified through automatic parsing of OpenSLR transcripts.

---

📚 Source

This dataset is based on the OpenSLR Nepali ASR corpus, distributed under the Creative Commons Attribution-ShareAlike 4.0 International License.

Source: https://www.openslr.org/54

---

⚖️ License

Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0)

This dataset inherits its license from the original OpenSLR Nepali ASR release. You are free to share and adapt the material for any purpose, even commercially, under the following terms:

• Attribution — You must give appropriate credit to the original OpenSLR dataset.
• ShareAlike — If you remix, transform, or build upon the material, you must distribute your contributions under the same license.

See the full license text here.

---

🧑‍💻 Acknowledgements

• OpenSLR contributors for curating and publishing the original Nepali ASR corpus.
• This repackaged dataset was created to simplify Whisper and multilingual ASR fine-tuning workflows.

---

🚀 Citation

If you use this dataset, please cite both OpenSLR and this preprocessed release.

@misc{openslr_nepali_asr,
  title = {Nepali ASR Corpus (OpenSLR #114)},
  author = {OpenSLR Contributors},
  year = {2022},
  url = {https://www.openslr.org/114},
  note = {Licensed under CC BY-SA 4.0}
}

@misc{nepali_asr_flac_prep,
  title = {Nepali ASR (FLAC) - Whisper Fine-Tuning Ready Dataset},
  author = {SAGEA},
  year = {2025},
  note = {Processed and released on Hugging Face Datasets}
}

---

✅ Summary

• 🌐 Language: Nepali (ne)
• 🔊 Format: .flac (16 kHz)
• 🧾 License: CC BY-SA 4.0
• 🧠 Task: Automatic Speech Recognition (ASR)
• ⚡ Ready for: Whisper, Wav2Vec2, Conformer, SpeechT5

---

Prepared and cleaned for research and educational purposes.