Separating the source information in repetition-dependent music and enhancing it by real-time digital audio processing

The repetitive patterns and rhythmic components in the music signal give enough clues about its structure. Using this insight, a repetition-based source separation technique was examined in our earlier work. A set of windowing techniques were examined to optimally separate an acoustic mixture of foreground vocal signal and the repetitive background musical accompaniments, into respective constituents. In this paper, we compare two approaches and their results, for achieving the music-source separation. First, a software-based approach is implemented, by finding the repetition-period from the beat-spectrum derived using the signal processing methods short-time Fourier transform and autocorrelation. The segments having repetitive patterns are identified, and then by taking the element-wise median for these segments, a normalization-based background model is derived from the music-speech mixture. In the second hardware-based approach, the digital signal processing (DSP) capabilities of a PSoC device are explored, for achieving the same objective. Experiments conducted in the first approach using different windowing techniques, indicate better results for the foreground separation, with Blackman window. But, for the background separation, use of Hamming window gives better results. Use of Kaiser window also gives good background-separation. In the second approach using DSP of PSoC devices, the use of High-Pass Filters gives better background music separation by suppressing the foreground vocals, and thereby giving a pleasant background music output. In both the experiments, the effects of digital audio processing are taken into account by use of post-processing, in the overall source-separation problem.