Vocal Isolation from Audio Tracks Using Neural Networks

Vocal Isolation from Audio Tracks Using Neural Networks

Ahish Bhat1, Sadroop G Prasad2, Arjun Chagalmarri3 , Vivek M Rao4

1Ahish Bhat, Dept of CS&E The National Institute of Engineering, Mysuru

2Sadroop G Prasad, Dept of CS&E The National Institute of Engineering, Mysuru

3Arjun Chagalmarri, Dept of IS&E The National Institute of Engineering, Mysuru

4Vivek M Rao, Dept of IS&E The National Institute of Engineering, Mysuru
5Dr. Shabana Sultana, Professor, Dept of CS&E The National Institute of Engineering, Mysuru
6Mrs. Soujanya K V, Assistant Professor, Dept of CS&E The National Institute of Engineering, Mysuru

Abstract - This paper presents a system for musical source separation using deep neural networks, specifically focusing on the U-Net architecture. The system processes audio data to separate individual components such as vocals and instrumental tracks from mixed music sources. We detail the implementation framework, preprocessing techniques, and post-processing methods used to achieve high-quality audio separation. Experimental results demonstrate the effectiveness of our approach, achieving Signal Distortion Ratio (SDR) values between 6.2-6.8 and Signal to Interference Ratio (SIR) values between 13-15 for two-stem separation tasks. We also compare our approach with state-of-the-art systems including Spleeter, Open-Unmix, and Demucs to provide comprehensive benchmarks for music source separation techniques.

Key Words:  Audio, Music. Music Source Separation, Deep Neural Networks, Wave U-Net Architecture, Audio Signal Processing, Vocal-Instrumental Separation