/odcnn

implementation of improved musical onset detection with cnn

Primary LanguagePython

Musical Onset Detection with Convolutional Neural Networks

Summary

  • model architecture: simple convolutional neural network

  • prediction: probability of onset

Description

This repository is implementation of Jan Schlüter and Sebastian Böck's "IMPROVED MUSICAL ONSET DETECTION WITH CONVOLUTIONAL NEURAL NETWORKS"

ABSTRACT:

Musical onset detection is one of the most elementary tasks in music analysis, but still only solved imperfectly for polyphonic music signals. Interpreted as a computer vision problem in spectrograms, Convolutional Neural Networks (CNNs) seem to be an ideal fit. On a dataset of about 100 minutes of music with 26k annotated onsets, we show that CNNs outperform the previous state-of-the-art while requiring less manual preprocessing. Investigating their inner workings, we find two key advantages over hand-designed methods: Using separate detectors for percussive and harmonic onsets, and combining results from many minor variations of the same scheme. The results suggest that even for well-understood signal processing tasks, machine learning can be superior to knowledge engineering.

Using PyTorch, implemented musical onset detectioner and apply it to Taikosanjiro which is sounds game like taiko-no-tatsujin

Usage

Requirement

  • git-lfs (for downloading large dataset)
  • python3
  • pytorch
  • soundfile
  • librosa
  • tqdm

Install Pytorch

you can download from here

Running Sample

  • install requirement
$ pip install -r reqirement.txt
  • before training, prepare audio dataset. (and pickle data.)
$ python music_processor.py reduced
$ python music_processor.py test
  • then, train model.
$ python train.py don
$ python train.py ka
  • then predict onset probability with trained model.
$ python infer.py don
$ python infer.py ka
  • and create notes and corresponding music from inferred data.
$ python synthesyze.py

Notice

  • train_reduced in this repo is too small for training, because of limitation of uploadable file size of git, I wasn't able to upload enough size training data.

  • if you want to train model with larger data, you could download audio data and corresponding notes here.

References