Evaluasi ICA dan NMF pada Pemisahan Sinyal Audio Menggunakan BSS Metrics dan MFCC

Abstract

Source separation is a crucial challenge in audio signal processing, particularly for stereo data. This study compares the performance of Independent Component Analysis (ICA) and Non-negative Matrix Factorization (NMF) in separating mixed audio signals. ICA operates directly on stereo signals, while NMF is applied to mono versions derived from stereo mixtures. Three pairs of audio data with diverse natural sound combinations were used. Evaluation metrics include Blind Source Separation indicators (SDR, SIR, SAR), spectral similarity based on Mel-Frequency Cepstral Coefficients (MFCC), and robustness tests by adding noise at 10 dB and 5 dB SNR levels. The results show that ICA consistently yields higher SDR and SIR scores and lower Euclidean distances in MFCC compared to NMF. In contrast, NMF performs poorly due to its mono-only limitation and inability to exploit spatial information. This study highlights ICA's superiority in separation accuracy and noise robustness, and emphasizes the importance of spectral analysis as a complementary evaluation method.