Comparative Evaluation of Fine-Tuned Hybrid Transformer and Band-Split Recurrent Neural Networks for Music Source Separation

In recent advancements within the field of music source separation (MSS), state-of-the-art models such as Hybrid Transformer Demucs (HT Demucs) and Band-Split Recurrent Neural Networks (BSRNN) have been at the forefront. Although the pre-trained HT Demucs model is capable of separating six sources—drums, bass, guitar, piano, vocals, and others—it underperforms on guitar, piano, and other sources compared to its performance on bass, drums, and vocals, as measured by the utterance-level Signal-to-Distortion Ratio (uSDR) metric. To date, there has been no evaluation of the BSRNN model’s ability to separate these six sources. This paper seeks to address this gap by investigating and comparing the performance of the BSRNN and HT Demucs models for six-source separation. Using the MoisesDB dataset, both models were developed and fine-tuned for this task. Their performance was then evaluated to identify the superior model for six-source separation. Experimental results reveal that the fine-tuned HT Demucs model surpasses the BSRNN model, achieving average uSDR and cSDR scores of 6.26 dB and 5.88 dB, respectively, compared to 5.52 dB and 5.38 dB for the BSRNN model. Moreover, the fine-tuned HT Demucs model outperforms its pre-trained counterpart on piano and other sources by 1 dB and 0.3 dB, respectively.