license: cc-by-sa-4.0
dataset_info:
- config_name: original
features:
- name: utterance_id
dtype: string
- name: speaker_id
dtype: string
- name: utterance
dtype:
audio:
sampling_rate: 16000
- name: transcription
dtype: string
- name: num_frames
dtype: int32
splits:
- name: train
num_bytes: 40925646
num_examples: 157905
download_size: 9340083067
dataset_size: 40925646
- config_name: cleaned
features:
- name: utterance_id
dtype: string
- name: speaker_id
dtype: string
- name: utterance
dtype:
audio:
sampling_rate: 16000
- name: transcription
dtype: string
- name: num_frames
dtype: int32
splits:
- name: train
num_bytes: 40925646
num_examples: 157905
download_size: 5978669282
dataset_size: 40925646
Dataset Card for OpenSLR Nepali Large ASR Cleaned
Table of Contents
Dataset Description
- Homepage: Original OpenSLR Large Nepali ASR Dataset link
- Repository: [Needs More Information]
- Paper: [Needs More Information]
- Leaderboard: [Needs More Information]
- Point of Contact: Sagar Sapkota
Dataset Summary
This data set contains transcribed audio data for Nepali. The data set consists of flac files, and a TSV file. The file utt_spk_text.tsv contains a FileID, anonymized UserID and the transcription of audio in the file. The data set has been manually quality-checked, but there might still be errors.
The audio files are sampled at a rate of 16KHz, and leading and trailing silences are trimmed using torchaudio's voice activity detection.
For your reference, following was the function applied on each of the original openslr utterances.
import torchaudio
SAMPLING_RATE = 16000
def process_audio_file(orig_path, new_path):
"""Read and process file in `orig_path` and save it to `new_path`"""
waveform, sampling_rate = torchaudio.load(orig_path)
if sampling_rate != SAMPLING_RATE:
waveform = torchaudio.functional.resample(waveform, sampling_rate, SAMPLING_RATE)
# trim end silences with Voice Activity Detection
waveform = torchaudio.functional.vad(waveform, sample_rate=SAMPLING_RATE)
torchaudio.save(new_path, waveform, sample_rate=SAMPLING_RATE)
How to use?
There are two configurations for the data: one to download the original data and the other to download the preprocessed data as described above.
- First, to download the original dataset with HuggingFace's Dataset API:
from datasets import load_dataset
dataset = load_dataset("spktsagar/openslr-nepali-asr-cleaned", name="original", split='train')
- To download the preprocessed dataset:
from datasets import load_dataset
dataset = load_dataset("spktsagar/openslr-nepali-asr-cleaned", name="cleaned", split='train')
Supported Tasks and Leaderboards
automatic-speech-recognition
: The dataset can be used to train a model for Automatic Speech Recognition.
Languages
Nepali
Dataset Structure
Data Instances
{
'utterance_id': 'e1c4d414df',
'speaker_id': '09da0',
'utterance': {
'path': '/root/.cache/huggingface/datasets/downloads/extracted/e3cf9a618900289ecfd4a65356633d7438317f71c500cbed122960ab908e1e8a/cleaned/asr_nepali/data/e1/e1c4d414df.flac',
'array': array([-0.00192261, -0.00204468, -0.00158691, ..., 0.00323486, 0.00256348, 0.00262451], dtype=float32),
'sampling_rate': 16000
},
'transcription': '२००५ मा बिते',
'num_frames': 42300
}
Data Fields
- utterance_id: a string identifying the utterances
- speaker_id: obfuscated unique id of the speaker whose utterances is in the current instance
- utterance:
- path: path to the utterance .flac file
- array: numpy array of the utterance
- sampling_rate: sample rate of the utterance
- transcription: Nepali text which spoken in the utterance
- num_frames: length of waveform array
Data Splits
The dataset is not split. The consumer should split it as per their requirements.