AI Sounds Detection Dataset

Description

This dataset contains audio clips designed for training Convolutional Neural Networks (CNN) to detect mechanical and household sounds.

Current Status

All audio files have been preprocessed and converted to 22050 Hz. The dataset is ready for use in training audio classification models.

Dataset Summary

• Total files in Sounds/: 5,987
• Total spectrogram files in Sounds/Spectrograms_Dataset/: 1,588
• Preprocessed audio files by class:
  • drilling/: 806
  • keyboard/: 94
  • vacuum/: 247
  • washing/: 444
• Raw source audio files: 68

Demo Application

Test the model trained on this dataset directly in the browser: AI Sounds Detection App

Classes

The dataset currently includes the following categories:

• drilling
• washing
• keyboard
• vacuum

Dataset Structure

The dataset is organized by sound class with the following folder structure:

• drilling/ - Drilling machine sounds
• keyboard/ - Keyboard typing sounds
  • fixed/ - Preprocessed and cleaned keyboard audio files
• vacuum/ - Vacuum cleaner sounds
  • fixed_vacuum/ - Preprocessed and cleaned vacuum audio files
• washing/ - Washing machine sounds
  • fixed_washing/ - Preprocessed and cleaned washing machine audio files
• raw/ - Original raw audio recordings that were collected before preprocessing
  • raw_keyboard/ - Raw keyboard recordings
  • raw_vacuum/ - Raw vacuum recordings
  • raw_washing/ - Raw washing machine recordings

Spectrogram Conversion

The repository now includes a ready-to-use spectrogram dataset in Sounds/Spectrograms_Dataset/.

• Sounds/Spectrograms_Dataset/drilling/ - 806 spectrogram images derived from drilling audio
• Sounds/Spectrograms_Dataset/keyboard/ - 93 spectrogram images derived from keyboard audio
• Sounds/Spectrograms_Dataset/vacuum/ - 246 spectrogram images derived from vacuum audio
• Sounds/Spectrograms_Dataset/washing/ - 443 spectrogram images derived from washing audio

These spectrograms are already converted and organized by class, making the dataset suitable for image-based audio classification workflows and CNN architectures trained on visual audio representations.

File Organization

• fixed/ folders contain preprocessed, cleaned, and normalized audio files ready for model training.
• raw/ folders contain original, unprocessed audio files that require preprocessing.
• Spectrograms_Dataset/ contains the class-specific spectrogram images generated from the cleaned audio files.

Notes

The improved dataset now includes both raw recordings and processed versions along with spectrogram representations, which provides a complete pipeline from raw audio collection to machine learning-ready input data.