A Unified Framework for Speech Dereverberation, Separation, and Denoising in Single-Channel Reverberant Noisy Mixtures

This paper presents a three-stage framework for single-channel speech separation and denoising in reverberant environments. The proposed system first applies Robust Principal Component Analysis (RPCA) to dereverberate the input mixture by decomposing the spectrogram into a low-rank component representing reverberation and a sparse component containing the dereverberated speech. This process also performs partial denoising, resulting in a cleaner mixture signal. In the second stage, we perform speaker separation using Sparse Non-negative Matrix Factorization (SNMF) with pre-trained speaker-specific dictionaries. The dereverberated mixture is factorized into speaker-structured basis components, allowing accurate reconstruction of individual speaker spectrograms. In the final stage, each separated spectrogram undergoes speaker-specific denoising through a second SNMF pass using clean training data. An Ideal Ratio Mask (IRM) is computed from the SNMF estimates to further suppress residual interference. Experimental evaluation confirms that the proposed method effectively addresses both reverberation and noise while achieving precise speaker separation, making it a promising approach for real-world speech enhancement applications.