My GitHub portfolio reports that Lyra achieves 94% precision on the OMAPS dataset, using its CNN-BiLSTM-Attention architecture over CQT spectrograms.
I built Lyra to convert audio into MIDI using deep learning, connecting my interest in music production with machine learning.
I represented audio as CQT spectrograms for the transcription task.
- I tried three model families: a simple CNN, a CNN-BiLSTM-Attention model, and a hybrid onset/frame model.
To handle how sparse notes are, I trained with a positively weighted binary cross-entropy loss using pos_weight 15.
I tracked experiments in Weights & Biases. No performance results or evaluation metrics were provided.