nn_audio

Audio Classifier

• Takes a name of a dataset in the form below as input

dataset___
          ├── class#1
          │   ├── file11.wav
          │   └── file12.wav
          ├── class#2
          │   ├── file21.wav
          │   ├── file22.wav
          │   └── file23.wav
          ├── class#3
          │   ├── file31.wav
          │   ├── file32.wav
          │   └── file33.wav
                 ...

• Uses the class names (i.e. class#1) as labels names
• Uses 10% of the data for testing and prints evaluation results.
• Returns an keras model and (optional) the data in a json file.

Requirements

• Python 3(recomended >=3.6)
• Pip (For installing other dependencies

Instalation

• Linux:
  • $ git clone https://github.com/Haki-Malai/nn-audio
  • $ cd nn-audio
  • $ sudo bash install.sh (Adds 'audio_model' as command. If it does not work try running 'source ~.profile')
• Windows:
  • Rename audio_model to audio_model.py

Usage

• Linux:
  • $ audio_model --dataset_dir=dataset (For example. dataset_dir is pwd for default but i don't think it is able to run this way as the script will read itself in the directory )
• Windows:
  • $ py audio_model.py --dataset_dir=dataset
• All options:
  • --sample_rate (default=44100)
  • --min_track_duration (default=10)
  • --min_array_length (default=1000000)
  • --from_json (default=False)
  • --json_save (default=False)
  • --output_json (default="data.json")
  • --json_path (default="data.json")
  • --model_name (default="model")
  • --test_size (default=0.3)
  • --epochs (default=50)
  • --activation (default='relu')
  • --dropout (default='True')
  • --num_segments (default=5)

A lot of data-preperation code and ml training is based on this channel's tutorial, make sure you subscribe to him if you are interested in Machine Learning and Audio compinations