Notice - A new project has been launched, which also contains this work. Please visit omnizart.
This is a Automatic Music Transcription (AMT) project, aim to deal with Multi-pitch Estimation (MPE) problem, which has been a long-lasting and still a challenging problem. For the transcription, we leverage the state-of-the-art image semantic segmentation neural network and attention mechanism for transcribing piano solo, and also multi-instrument performances.
The dataset used is MAPS and MusicNet, which the first one is a solo-piano performance collection, and the second is a multi-instrument performance collection. On both dataset, we achieved the state-of-the-art results on MPE (Multi-Pitch Estimation) case frame-wisely, which on MAPS we achieved F-score 86.73%, and on MusicNet we achieved F-score 73.70%.
This work was done based on our prior work of repo1, repo2. For more about our works, please meet our website.
For whom would interested in more technical details, the original paper is here.
The most straight forward way to enjoy our project is to use our colab. Just press the start button cell by cell, and you cant get the final output midi file of the given piano clip.
A more technical way is to download this repository by executing git clone https://github.com/BreezeWhite/Music-Transcription-with-Semantic-Segmentation.git, and then enter scripts folder, modify transcribe_audio.sh, then run the script.
One of the main topic in AMT is to transcribe a given raw audio file into symbolic form, that is transformation from wav to midi. And our work is the middle stage of this final goal, which we first transcribe the audio into what we called "frame level" domain. This means we split time into frames, and the length of each frame is 88, corresponding to the piano rolls. And then make prediction of which key is in activation.
Here is an example output:
The top row is the predicted piano roll, and the bottom row is the original label. Colors blue, green, and red represent true-positive, false-positive, and false-negative respectively.
We used semantic segmentation model for transcription, which is also widely used in the field of image processing. This model is originally improved from DeepLabV3+, and further combined with U-net architecture and focal loss, Illustrated as below:
To install the requirements, enter the following command:
pip install -r requirements.txt
Download weights of the check points:
git lfs fetch
For a quick example usage, you can enter scripts folder and check the code in the script to see how to use the python code.
-
Download dataset from the official website of MAPS and MusicNet.
-
cd scripts -
Modify the content of generate_feature.sh
-
Run the generate_feature.sh script
There are some cases for training, by defining different input feature type and output cases. For quick start, please refer to scripts/train_model.sh
For input, you can either choose using HCFP or CFP representation, depending on your settings of pre-processed feature.
For output, you can choose to train on MPE mode or multi-instrument mode, if you are using MusicNet for training. If you are using MAPS for training, then you can only train on MPE mode.
To train the model on MusicNet, run the command:
python3 TrainModel.py MusicNet \
<output/model/name>
--dataset-path <path/to/extracted/feature> \
The default case will train on MPE by using CFP features. You can train on multi-instrument mode by adding --multi-instruments flag, or change to use HCFP feature by adding --use-harmonic flag.
There are also some cases you can specify to accelerate the progress. Specify --use-ram to load the whole features into the ram if your ram is big enough (at least 64 GB, suggested >100 GB).
To validate the execution of the training command, you can also specify less epochs and steps by adding -e 1 -s 500.
And to continue train on a pre-trained model, add --input-model <path/to/pre-trained/model>.
To predict and evaluate the scores with label, run the command:
python3 Evaluation.py frame \
--feature-path <path/to/generated/feature> \
--model-path <path/to/trained/model> \
--pred-save-path <path/to/store/predictions> \
You can check out scripts/evaluate_with_pred.sh and scripts/pred_and_evaluate.sh for example use.
To transcribe on a single song, run the command:
python3 SingleSongTest.py \
--input-audio <input/audio>
--model-path <path/to/pre-trained/model>
There will be an output file under the same path named pred.hdf, which contains the prediction of the given audio.
To get the predicted midi, add --to-midi <path/to/save/midi> flag. The midi will be stored at the given path.
There is also an example script in scripts folder called transcribe_audio.sh