metadata

dataset_info:
  features:
    - name: audio
      dtype: audio
    - name: transcription
      dtype: string
  splits:
    - name: train
      num_bytes: 5844021677.902
      num_examples: 7481
    - name: test
      num_bytes: 526633107
      num_examples: 726
  download_size: 5452408390
  dataset_size: 6370654784.902
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*
license: mit
task_categories:
  - automatic-speech-recognition
language:
  - ne

Nepali Speech-to-Text Dataset

This repository contains a dataset for Automatic Speech Recognition (ASR) in the Nepali language. The dataset is designed for supervised learning tasks and includes audio files along with their corresponding transcriptions. The audio samples have been collected from various open-source platforms and other publicly available sources on the internet.

Each audio file has an average length of 15 seconds and has been converted into a consistent WAV format for ease of processing.

Dataset Structure

The dataset is split into training and testing sets:

Training Data: Contains a diverse set of Nepali speech samples from multiple sources.
Testing Data: Includes the Fleurs test data for Nepali to ensure evaluation consistency.

Audio Data Overview

The total dataset contains approximately 22.87 hours of audio. Below is the breakdown of the dataset:

Table 1: Audio data length from different sources

Dataset	Audio size (Hrs)
Common Voice 20	1.71
Google Fleurs	14.38
OpenSLR 143	1.24
OpenSLR 43	2.80
Nepali Parliament Audio	2.74
Total	22.87

The dataset includes high variability in terms of speakers (age groups, genders), noisy environments, different dialects, and various acoustic conditions, making it robust for ASR training.

Important Notes

The dataset is in raw form, meaning preprocessing and other corrections may be required before training an ASR model.
The transcriptions have been obtained from open datasets and may contain errors or inconsistencies that need to be addressed during data preparation.

References

Conneau, A., Ma, M., Khanuja, S., Zhang, Y., Axelrod, V., Dalmia, S., Riesa, J., Rivera, C., & Bapna, A. (2022).
FLEURS: Few-shot Learning Evaluation of Universal Representations of Speech.
arXiv preprint arXiv:2205.12446. Link
Sodimana, K., Pipatsrisawat, K., Ha, L., Jansche, M., Kjartansson, O., De Silva, P., & Sarin, S. (2018).
A Step-by-Step Process for Building TTS Voices Using Open Source Data and Framework for Bangla, Javanese, Khmer, Nepali, Sinhala, and Sundanese.
Proc. The 6th Intl. Workshop on Spoken Language Technologies for Under-Resourced Languages (SLTU), Gurugram, India, pp. 66-70.
DOI Link
Khadka, S., G.C., R., Paudel, P., Shah, R., & Joshi, B. (2023).
Nepali Text-to-Speech Synthesis using Tacotron2 for Melspectrogram Generation.
SIGUL 2023, 2nd Annual Meeting of the Special Interest Group on Under-resourced Languages: a Satellite Workshop of Interspeech 2023.
Ardila, R., Branson, M., Davis, K., Henretty, M., Kohler, M., Meyer, J., Morais, R., Saunders, L., Tyers, F. M., & Weber, G. (2020).
Common Voice: A Massively-Multilingual Speech Corpus.
Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pp. 4211-4215.