gam30/Nepali-asr-train-val

Nepali ASR Train and Validation Set (Clean & Noisy)

This dataset contains approximately 19.41 hours of total Nepali speech data sourced from OpenSLR, organized into an 80:20 train and validation split. To improve model robustness in real-world scenarios, 40% of the samples in both splits have been augmented with background noise (crowd, traffic, construction, and wind). The remaining 60% consists of clean audio.

Dataset Overview

Property	Value
Language	Nepali
Source	OpenSLR
Total Samples	19,775
Noise Type	Synthetic environmental noise (crowd, traffic, construction, wind)
Noise Coverage	40% augmented, 60% clean
Audio Format	WAV
Sample Rate	16,000 Hz (16 kHz)

Dataset Features

Each sample in the dataset contains:

• audio (Audio): The audio waveform data

  • Sample Rate: 16,000 Hz (16 kHz)
  • Channels: 1 (Mono)

• text (String): Full Nepali transcription of the speech

• duration (Float): Audio duration in seconds

Noise Characteristics

40% of the audio samples contain synthetic environmental noise mixed with the clean Nepali speech:

• Noise Sources:
  • 🏢 Crowd noise (background conversations, ambient chatter)
  • 🚗 Traffic noise (vehicle engines, horns, road sounds)
  • 🏗️ Construction noise (machinery, tools, equipment)
  • 💨 Wind noise (outdoor wind, air movements)

Loading the Dataset

Using HuggingFace datasets library

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("gam30/Nepali-asr-train-val")

# Access samples from the train split
sample = dataset['train'][0]
print(f"Transcription: {sample['text']}")
print(f"Duration: {sample['duration']}s")

# Iterate through dataset
for sample in dataset['train']:
    text = sample['text']
    duration = sample['duration']
    print(f"{duration}s - {text[:50]}...")

With Audio Feature

from datasets import load_dataset, Audio

dataset = load_dataset("gam30/Nepali-asr-train-val")
dataset = dataset.cast_column("audio", Audio(sampling_rate=16000))

# Access audio data
sample = dataset['train'][0]
print(f"Array shape: {sample['audio']['array'].shape}")
print(f"Sampling rate: {sample['audio']['sampling_rate']} Hz")  # Will be 16000

Working with Audio Files

from datasets import load_dataset
import librosa

dataset = load_dataset("gam30/Nepali-asr-train-val")

# Process first 5 samples
for sample in dataset['train'][:5]:
    # Load audio using librosa
    audio_path = sample['audio']['path']
    y, sr = librosa.load(audio_path, sr=None)
    
    print(f"File: {audio_path}")
    print(f"Sampling rate: {sr} Hz")
    print(f"Duration: {len(y) / sr:.2f}s")
    print(f"Text: {sample['text'][:60]}...")
    print()

Use Cases

This dataset is suitable for:

1. Robust ASR Model Training - Training models on noisy speech
2. Noise Robustness Testing - Evaluating ASR systems on noisy conditions
3. Domain Adaptation - Fine-tuning pre-trained models on Nepali
4. Speech Enhancement Research - Testing denoising techniques

Dataset Statistics

• Total Samples: 19,775
• Total Audio Duration: ~19.41 hours
• Train Split: 15,820 samples (~15.59 hours)
• Validation Split: 3,955 samples (~3.82 hours)
• Sample Rate: 16,000 Hz (16 kHz Mono) - standard for ASR tasks

Citation

If you use this dataset in your research, please cite:

@dataset{nepali_asr_noisy_2024,
  title={Nepali ASR Train and Validation noisy set},
  author={sangam},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/gam30/Nepali-asr-train-val}
}

Quality Assurance

• ✓ All transcriptions in UTF-8 Unicode format
• ✓ Duration metadata computed and validated
• ✓ Audio verified at 16,000Hz mono

Dataset Structure

gam30/Nepali-asr-train-val
├── train/
│   ├── audio (Audio)
│   ├── text (String) 
│   └── duration (Float)
└── val/
    ├── audio (Audio)
    ├── text (String) 
    └── duration (Float)

Support & Issues

For questions or issues with the dataset:

1. Check the Hugging Face community discussions
2. Open an issue on the dataset repository

---

Dataset ID: gam30/Nepali-asr-train-val
Last Updated: 2026