Abstract
Blind Source Separation (BSS) often introduces "musical noise"—distorting artifacts caused by discontinuous Time-Frequency (TF) masking. This paper proposes a time-domain post-processing method to mitigate these effects. By employing $l_1$ regularization and the fast Split Bregman optimization method to estimate sparse filters, the proposed approach reconstructs separated signals with significantly higher audio quality. Experimental results demonstrate substantial improvements in Signal-to-Artifact Ratio (SAR) and subjective listener preference compared to traditional masking techniques.
Introduction
In Blind Source Separation (BSS), the goal is to recover individual sound sources from a mixture recorded by multiple microphones. A popular approach is Time-Frequency (TF) binary masking, which assumes that at any specific moment and frequency, one source is dominant. However, this "winner-take-all" approach is physically unnatural and creates jagged "holes" in the spectrogram, resulting in a chirping, watery distortion known as musical noise. This is especially problematic in music or environments with many sources where the "one source per TF point" assumption fails.
Methodology
The authors propose a two-stage improvement to address these artifacts:
- Refined TF Masking: A "confidence coefficient" is introduced to the initial masking process. Instead of blindly assigning every point to a cluster, the algorithm filters out "fuzzy" points—those that sit on the edge between potential sources—to reduce initial errors.
- Time-Domain Sparse Filters: This is the core innovation. Instead of remaining in the frequency domain, the method moves to the time domain. It seeks a pair of sparse filters ($u$) that, when applied to the original microphone mixtures ($x$), reconstruct the masked output ($y$) while minimizing noise.
- Split Bregman Optimization: Solving for these sparse filters involves $l_1$ regularization. To make this computationally efficient, the authors employ the Split Bregman method, which allows them to solve the complex optimization problem in approximately 5 seconds per channel—fast enough for practical use.
Experimental Setup
The approach was evaluated using:
- Synthetic Mixtures: Created using measured binaural Room Impulse Responses (RIRs) with varying reverberation times (anechoic, 150ms, and 580ms).
- Real-World Room Recordings: Conducted in a 4.5m x 3.6m room with a 130ms reverberation time. Tests included mixtures of 2 and 4 sources (male/female speech and music) recorded at an 8000Hz sampling rate.
Key Results and Discussion
The proposed method was compared against standard binary masks (BM), Sigmoid masks (SM), and Bayesian masks (BYM).
Objective Quality
In room recordings, the proposed method achieved a PESQ score of 2.40, significantly outperforming the baseline binary masks (2.24) and Sigmoid masks (2.17). The Signal-to-Artifact Ratio (SAR) saw massive gains, jumping from ~9.37 dB in the baseline to 19.05 dB with the new method.
As shown in the table above, while Bayesian masks may occasionally offer higher interference rejection (SIR), the proposed method provides superior overall signal reconstruction (SDR) and artifact reduction (SAR).
Subjective Preference
In listening tests, human subjects showed a clear preference for the proposed method's output. The reduction in "musical noise" was particularly noticeable in music tracks.
Subjects preferred the proposed method over binary masks 94% of the time for speech and 99% of the time for music. Even when compared to more advanced Sigmoid and Bayesian masks, the proposed method maintained a strong lead or parity in most categories.
Conclusion
By enforcing $l_1$ regularization on the filters, the algorithm naturally seeks the "simplest" explanation for the sound, effectively acting as a de-reverberation tool. By using the frequency-domain output as a target for a time-domain filter, the authors achieve the separation power of masking with the signal smoothness of filtering. Future work will focus on further optimizing the Split Bregman implementation for real-time, low-latency applications.
Summary updated