Abstract
Traditional two-microphone noise reduction algorithms to deal with highly nonstationary directional noises generally use the direction of arrival or phase difference information. The performance of these algorithms deteriorate when diffuse noises coexist with nonstationary directional noises in realistic adverse environments. In this paper, we present a two-channel noise reduction algorithm using a spatial information-based speech estimator and a spatial-information-controlled soft-decision noise estimator to improve the noise reduction performance in realistic non-stationary noisy environments. A target presence probability estimator based on Bayes rules using both phase difference and magnitude squared coherence is proposed for soft-decision of noise estimation, so that they can share complementary advantages when both directional noises and diffuse noises are present. Performances of the proposed two-microphone noise reduction algorithm are evaluated by noise reduction, log-spectral distance (LSD) and word recognition rate (WRR) of a distant-talking ASR system in a real room's noisy environment. Experimental results show that the proposed algorithm achieves better noises suppression without further distorting the desired signal components over the comparative dual-channel noise reduction algorithms.
Introduction
Dual-microphone speech enhancement is critical for modern communication devices, yet traditional methods often struggle in "messy" acoustic environments where directional interferences (like a nearby TV) and diffuse background noise (like office chatter) coexist. Most existing algorithms rely solely on Phase Difference (PD) or Direction of Arrival (DOA) to isolate the target speaker. However, in reverberant or diffuse fields, phase information becomes unreliable, and aggressive signal subtraction can lead to "speech leakage," where the target voice is accidentally suppressed along with the noise.
This paper proposes a robust alternative that combines two distinct spatial cues—PD and Magnitude Squared Coherence (MSC)—within a Bayesian framework to improve speech presence detection and noise estimation accuracy.
Proposed Methodology
The proposed system architecture, shown in Figure 1, avoids the traditional "blocking matrix" subtraction approach in favor of a probabilistic spectral amplitude estimation framework.
Key Components:
- Dual-Cue Target Presence Probability (TPP): The system calculates the probability that a specific time-frequency bin contains target speech by combining PD (effective for directional noise) and MSC (effective for diffuse noise) using Bayes' rule.
- Spatial-Information-Controlled Soft-Decision (SCSD): Instead of hard subtraction, the TPP is used to update the noise model. The noise estimate is updated only when the probability of speech presence is low, effectively preventing target speech from leaking into the noise model.
- Auditory Processing: To mimic human hearing and improve robustness, the system employs a bank of 40 Gammatone filters for frequency-domain smoothing.
- Modified LSA Estimator: A Log-Spectral Amplitude (LSA) estimator is used to apply the final gain, minimizing the distance between the estimated and clean speech spectra.
Experimental Results
The algorithm was evaluated using clean speech from the TIMIT database mixed with simulated and real-world noise. Simulations included a room with a 300ms reverberation time ($RT_{60}$) and non-stationary interferences at specific angles.
As illustrated in Figure 3, the proposed SCSD method (red line) tracks the true noise power (blue line) significantly more accurately than traditional methods like MCRA (dash-dot line) or PDSAE (dashed green line). While MCRA often underestimates non-stationary noise and PDSAE exhibits high variance, the SCSD approach provides a stable and precise estimate even during rapid changes in interference, such as music or secondary voices.
Quantitative Performance:
- Noise Suppression: The proposed method achieved superior Noise Reduction (NR) and lower Log-Spectral Distance (LSD) across all tested Signal-to-Noise Ratios (-5dB to 10dB).
- ASR Accuracy: In real-world office tests with interfering voices, the Word Recognition Rate (WRR) improved from a baseline of 46.1% (noisy) to 70.9% (proposed), outperforming the next best method (PDSAE) by approximately 6% absolute.
Conclusion
By leveraging the complementary strengths of phase difference and magnitude squared coherence, the proposed algorithm provides a robust solution for speech enhancement in complex acoustic environments. The soft-decision noise estimation strategy effectively suppresses non-stationary interferences while preserving the integrity of the target speech. Future work will explore automatic speaker tracking and further optimization for real-time deployment on low-power mobile hardware.
Summary updated