Implementing Transformer Architectures for Audio Source Separation

Audio source separation, utilized for speech denoising and music production, involves extracting individual sources from a mixed track. The project goal was to reduce training times for separation models with limited computing resources while improving accuracy, measured through signal-to-distortion ratio (SDR) and signal-to-interference ratio (SIR). We replaced the Bi-directional Long Short-Term Memory (BiLSTM) block with a transformer in the open-source Open-Unmix model, training it on the MUSDB18-HQ dataset. Ultimately, the transformer implementation trained faster than the unmodified Open-Unmix model while achieving a higher SIR but a lower SDR, demonstrating the potential of transformers in increasing efficiency and decreasing track bleed-through.