https://github.com/marionchaff/piano-transcription

Turn audio recordings into piano sheets
https://github.com/marionchaff/piano-transcription

audio-files audio-processing deep-learning midi midi-files music piano piano-sheet-music python spectrogram tensorflow

Last synced: about 2 months ago
JSON representation

Turn audio recordings into piano sheets

• Host: GitHub
• URL: https://github.com/marionchaff/piano-transcription
• Owner: MarionChaff
• Created: 2024-08-25T17:04:03.000Z (4 months ago)
• Default Branch: main
• Last Pushed: 2024-08-25T17:06:29.000Z (4 months ago)
• Last Synced: 2024-10-12T04:41:25.663Z (3 months ago)
• Topics: audio-files, audio-processing, deep-learning, midi, midi-files, music, piano, piano-sheet-music, python, spectrogram, tensorflow
• Language: Jupyter Notebook
• Homepage:
• Size: 105 KB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
# piano-transcription

Turn audio recordings into piano sheets

This repo contains two notebooks including (i) audio data preprocessing and (ii) the training of a model designed to recognize piano notes. It can be used to turn audio recordings into piano sheets.

**Database**

- Very short midi files with one or many notes played simultaneously (midi files contains information such as as musical notes, timing, velocity, pitch, and instrument data)

**X data pre-processing** (cf. file 1-data-preprocessing.ipynb)

- Each midi (.mid) is converted into audio (.wav) file

- Each audio file (.wav) is converted into a CQT spectrogram (.png)

- Each spectrogram (.png) is converted into a grey-scale 1d array and resized to a unique format (60 x 150)

**y data pre-processing** (cf. file 1-data-preprocessing.ipynb)

- The notes played are extracted from the midi (.mid) files and are one-hot encoded

**Model training** (cf. file 2-model-training.ipynb)

- Convolutional layers were used to extract spatial features from each spectogram

- Sigmoid activation and binary crossentropy loss functions were used to capture multiple notes played simultaneously

- Model accuracy reaches 0.96 on train, val and test sets