Skip to content

Code for YouTube series: Deep Learning for Audio Classification

License

Notifications You must be signed in to change notification settings

seth814/Audio-Classification

Repository files navigation

Audio-Classification (Kapre Version)

Pipeline for prototyping audio classification algorithms with TF 2.3

melspectrogram

YouTube

This series has been re-worked. There are new videos to support this repository. It is recommended to follow the new series.

https://www.youtube.com/playlist?list=PLhA3b2k8R3t0SYW_MhWkWS5fWg-BlYqWn

If you want to follow the old videos, restore to a previous commit.

git checkout 404f2a6f989cec3421e8217d71ef070f3593a84d

Environment

conda create -n audio python=3.7
activate audio
pip install -r requirements.txt

Jupyter Notebooks

Assuming you have ipykernel installed from your conda environment

ipython kernel install --user --name=audio

conda activate audio

jupyter-notebook

Audio Preprocessing

clean.py can be used to preview the signal envelope at a threshold to remove low magnitude data

When you uncomment split_wavs, a clean directory will be created with downsampled mono audio split by delta time

python clean.py

signal envelope

Training

Change model_type to: conv1d, conv2d, lstm

Sample rate and delta time should be the same from clean.py

python train.py

Plot History

Assuming you have ran all 3 models and saved the images into logs, check notebooks/Plot History.ipynb

history

notebooks/Confusion Matrix and ROC.ipynb

Confusion Matrix

conf_mat

Receiver Operating Characteristic

roc

Kapre

For computation of audio transforms from time to frequency domain on the fly

https://github.com/keunwoochoi/kapre
https://arxiv.org/pdf/1706.05781.pdf