autochord

Automatic Chord Recognition tools

About

autochord is:

✔ a Python library for automatic chord recognition (using TensorFlow)

✔ a Javascript app for visualization of chord transcriptions:

Library Usage

To install library, run:

$ pip install autochord

autochord provides a very simple API for performing chord recognition:

import autochord
autochord.recognize('audio.wav', lab_fn='chords.lab')
# This gives out a list of tuples in the format:
#  (chord start, chord end, chord name)
# e.g.
# [(0.0, 5.944308390022676, 'D:maj'),
#  (5.944308390022676, 7.476825396825397, 'C:maj'),
#  (7.476825396825397, 18.250884353741498, 'D:maj'),
#  (18.250884353741498, 19.736961451247165, 'C:maj')
#  ...
#  (160.49632653061224, 162.30748299319728, 'N')]

Under the hood autochord.recognize() runs the NNLS-Chroma VAMP plugin to extract chroma features from the audio, and feeds it to a Bi-LSTM-CRF model in TensorFlow to recognize the chords. Currently, the model can recognize 25 chord classes: the 12 major triads, 12 minor triads, and no-chord ('N').

OPTIONALLY, you may dump the chords in a .lab file by using the lab_fn parameter. The output file follows the MIREX chord label format.

Upon import autochord takes care of setting up the VAMP plugin and downloading the pre-trained chord recognition model.

The measured test accuracy of the TensorFlow model is 67.33%. That may be enough for some songs, but we can explore in the future how to further improve this.

App Usage

The app is pretty straightforward: you need to load a song, then you can upload a LAB file to visualize its chord labels. You may use the autochord Python library for generating this file. Optionally, you may load another LAB file for comparison (e.g. ground-truth labels, LAB file from another model's prediction).

Future Improvements

• Integrate everything into a full chord recognition app! For that we need to:
  • convert VAMP plugin to JS module
  • model conversion to TensorFlow.js (as of writing, some CRF operations are not supported by TFJS yet)
  • converting all other Python functions to JS equivalent
• Experimenting with other approaches to improve chord recognition accuracy