phoneme-informed-note-level-singing-transcription

A pretrained model for "A Phoneme-informed Neural Network Model for Note-level Singing Transcription", ICASSP 2023

Requirements

• torch==1.13.1
• torchaudio==0.13.1
• nnAudio==0.3.2
• mido==1.2.10
• mir_eval==0.7
• librosa==0.9.1
• numpy==1.23.4
• wquantiles==0.4

Or if you are using Poetry, you can install the dependencies by running

$ poetry install

Usage

$ python infer.py checkpoints/model.pt INPUT_FILE OUTPUT_FILE --bpm BPM_OF_INPUT_FILE --device DEVICE

• INPUT_FILE is the path to the input audio file.
• OUTPUT_FILE is the path to the output MIDI file. (If you do not give this argument, the default file name will be out.mid.)
• BPM_OF_INPUT_FILE is the BPM of the input audio file. (If you do not give this argument, the default value will be 120.)
• DEVICE is the device to run the model. (If you do not give this argument, the default device will be cuda:0 if available, otherwise cpu.)

Blog with Demo Examples

https://seyong92.github.io/phoneme-informed-transcription-blog/

Remarks

To pull the model checkpoint from the GitHub repository, Git LFS is needed.

For people who suffer for downloading the model checkpoint through Git LFS, I uploaded the model checkpoint in this link.