Abstract
This paper presents a novel post-filter for noise reduction. A subspace based noise estimation method is developed with the use of multiple statistical distributions to model the speech and noise. The signal-plus-noise subspace dimension is determined by maximizing the target speech presence probability in noisy frames, so as to estimate the noise power spectrum for post-filter design. Then, masking property is incorporated in the post-filter technique for residual noise shaping. Experimental results show that the proposed scheme outperforms the baseline systems in terms of various quality measurements of the enhanced speech.
Introduction
Microphone arrays utilize beamforming to spatially filter out noise from undesired directions. However, beamforming alone often leaves significant background noise, necessitating the use of post-filters. Traditional post-filters, such as the Zelinski or McCowan filters, frequently suffer from inaccurate noise estimation (often relying on Voice Activity Detection during pauses) and poor perceptual quality due to distracting "musical noise."
This paper introduces an advanced post-filter that combines subspace-based noise estimation with psychoacoustic masking. By adaptively modeling speech distributions and shaping residual noise to be "hidden" from the human ear, the proposed method significantly improves speech quality and noise reduction compared to traditional techniques.
Proposed Methodology
The proposed system architecture consists of two primary stages: adaptive subspace noise estimation and masking-property based post-filtering.
Adaptive Subspace Noise Estimation
Instead of relying on speech pauses, the system employs subspace decomposition to separate the signal into "signal-plus-noise" and "noise-only" subspaces. To improve robustness, the method utilizes a mixture of Gaussian, Laplacian, and Gamma distributions, adaptively selecting the optimal model for different frequency bins. The signal subspace dimension is determined by maximizing the speech presence probability, allowing for accurate noise estimation even during active speech segments.
Masking-Property Based Post-Filter
The second stage leverages psychoacoustics to improve perceptual quality. By calculating a masking threshold—the level below which sounds are inaudible to the human ear—the post-filter shapes the residual noise. Rather than attempting to eliminate all noise, which can lead to speech distortion, the filter ensures that leftover noise remains below the threshold, effectively making it inaudible to the listener.
Experimental Evaluation
The system was evaluated using the CMU Microphone Array Database with an 8-sensor linear array. Testing involved two noise types: steady "machine engine noise" and non-steady "people cheer noise," across a range of input Signal-to-Noise Ratios (SNR) from -10dB to 10dB.
Key Results
The proposed method was compared against standard beamforming and established post-filters (Zelinski and McCowan).
- Noise Estimation: The subspace method achieved a Noise Estimation Error (NEE) of 0.73–0.99, significantly lower than the 2.12–4.56 range typical of VAD-based methods.
- SNR Improvement: The method provided average SNR gains of 2.95dB for engine noise and 3.32dB for cheer noise over the best reference approach.
- Speech Quality: Significant improvements were observed in the Itakura-Saito (IS) distortion metric (23.9% improvement for engine noise) and the Log-Likelihood Ratio (LLR), indicating more natural-sounding enhanced speech.
Conclusion
By integrating adaptive subspace noise estimation with psychoacoustic masking, this research provides a robust solution for microphone array post-filtering. While the method introduces higher computational complexity due to eigenvalue decomposition, it offers superior perceptual quality and noise reduction compared to traditional MSE-based filters. Future work may focus on optimizing real-time performance and refining masking threshold estimation in highly reverberant environments.
Summary updated