Firoj112/highquality_nepali_female_asr

🗣️ Nepali Female Speech Dataset (SLR43 Subset)

📖 Dataset Summary

This dataset is a restructured and Hugging Face–ready version of the OpenSLR SLR43: Multi-speaker TTS Data for Nepali (ne-NP). It consists of high-quality speech recordings from Nepali female speakers, originally curated by Google LLC.

The dataset has been reformatted for Automatic Speech Recognition (ASR) and Speech-to-Text (STT) fine-tuning, maintaining full Nepali (Devanagari) transcriptions and consistent sampling at 16 kHz.

---

🧩 Dataset Structure

Field	Description
utt_id	Unique utterance identifier
speaker_id	Speaker label (here: nep_female)
sentence	Transcribed Nepali text (Devanagari script)
audio	Audio file (16kHz sampling rate)

Example

{
  "utt_id": "nep_0258_1139532706",
  "speaker_id": "nep_female",
  "sentence": "रेडियो नेपालमा कार्यरत रहँदा मास्टर रत्नदास प्रकाश निकै लोकप्रिय थिए",
  "audio": {
    "path": "wavs/nep_0258_1139532706.wav",
    "array": [...],
    "sampling_rate": 16000
  }
}

Splits

Split	# Examples	Size (bytes)
Train	1,857	~871 MB
Test	207	~98 MB
Total	2,064	~970 MB

---

🗂️ Dataset Organization

Original source:

ne_np_female.zip
├── line_index.tsv
└── wavs/
    ├── nep_0258_0119737288.wav
    ├── nep_0258_0461984530.wav
    ├── ...

Converted structure for ASR:

• Combined line_index.tsv and corresponding .wav files
• Added absolute audio paths and metadata fields (utt_id, speaker_id, sentence, audio)
• Split into train and test subsets

---

⚙️ Usage Example

from datasets import load_dataset

dataset = load_dataset("Firoj112/highquality_nepali_female_asr")
print(dataset["train"][0])

# For Whisper fine-tuning
from datasets import Audio
dataset = dataset.cast_column("audio", Audio(sampling_rate=16000))

---

📄 License

This dataset is distributed under the Attribution-ShareAlike 4.0 (CC BY-SA 4.0) license. Refer to the OpenSLR SLR43 page for full license terms.

---

📚 Citation

If you use this dataset in your research or fine-tuning, please cite the original publication:

@inproceedings{kjartansson-etal-tts-sltu2018,
  title = {{A Step-by-Step Process for Building TTS Voices Using Open Source Data and Framework for Bangla, Javanese, Khmer, Nepali, Sinhala, and Sundanese}},
  author = {Keshan Sodimana and Knot Pipatsrisawat and Linne Ha and Martin Jansche and Oddur Kjartansson and Pasindu De Silva and Supheakmungkol Sarin},
  booktitle = {Proc. The 6th Intl. Workshop on Spoken Language Technologies for Under-Resourced Languages (SLTU)},
  year  = {2018},
  address = {Gurugram, India},
  month = aug,
  pages = {66--70},
  URL   = {http://dx.doi.org/10.21437/SLTU.2018-14}
}

---

⚠️ Disclaimer

This dataset is a repackaged educational version of the original SLR43 Nepali female speech dataset. All rights remain with the original creators — © 2016–2018 Google LLC.

It is provided here for research, educational, and fine-tuning purposes only.

---

👩‍💻 Maintainer Notes

• Converted to Hugging Face–compatible ASR dataset format
• Audio normalized to 16kHz
• Text preserved in Devanagari script (Nepali)
• Pre-split into train/test for rapid experimentation

---

📦 Reformatted and prepared for educational use by Firoj112.