Willy030125/ambient_noise_audio

Ambient Noise Collection dataset

This dataset is useful for reducing ASR model Hallucinations (especially Whisper), by default Whisper often transcribing non-speech audio as hallucination transcriptions.<br> This dataset attempts to improve ASR model that have hallucinations on non-speech audio.

Collected several audio from source:

1. 1.https://www.kaggle.com/datasets/nafin59/hospital-ambient-noise
2. 2.https://www.kaggle.com/datasets/solorzano/ambient-noise
3. 3.https://www.kaggle.com/datasets/minsithu/audio-noise-dataset
4. 4.https://www.kaggle.com/datasets/ivanj0/audiodata
5. 5.https://huggingface.co/datasets/Myrtle/CAIMAN-ASR-BackgroundNoise

Dataset info:<br> num_rows: 288<br>

Format

Each example is a dictionary with the following fields:

{
  "path": "audio/ambient_noise_(1).wav",
  "audio": {
    "path": "audio/ambient_noise_(1).wav",
    "array": [...],
    "sampling_rate": 16000
  },
  "sentence": "<|nospeech|>"
}

Load dataset

Use HuggingFace datasets v2.18:

pip install datasets==2.18.0

Use HuggingFace datasets to load:

from datasets import load_dataset, Audio

try:
    dataset = load_dataset("Willy030125/ambient_noise_audio", trust_remote_code=True, encoding="utf-8", errors="ignore")
except:
    dataset = load_dataset("Willy030125/ambient_noise_audio", trust_remote_code=True)