Abstract
In this paper, we propose an improved noise power estimation method for speech enhancement under various noise environments based on Minimum Statistics (MS). Traditional MS algorithms use a fixed search window size for tracking the minimum value of the power spectrum, which is suboptimal across diverse noise conditions. Our proposed algorithm introduces environment awareness by using a Gaussian Mixture Model (GMM) to classify the noise type and adaptively adjust the search window size. Performance evaluation using the ITU-T P.862 Perceptual Evaluation of Speech Quality (PESQ) metric demonstrates that the proposed method yields superior results compared to the conventional MS method across various noise environments.
Introduction
Speech enhancement is critical for improving communication quality in noisy environments. A fundamental component of many enhancement systems is noise power estimation. The Minimum Statistics (MS) method is widely used because it can estimate noise power even during active speech by tracking the minimum of the smoothed power spectrum within a finite search window. However, the performance of MS is highly dependent on the choice of the search window size ($D$). A fixed window size often fails to balance the trade-off between tracking speed in non-stationary noise and stability in stationary noise.
Proposed Environment-Aware MS Algorithm
To address the limitations of a fixed search window, we propose an adaptive MS algorithm that adjusts its parameters based on the specific noise environment.
Noise Classification
The system first identifies the noise environment using a Gaussian Mixture Model (GMM). A 14-dimensional feature vector is extracted from each frame, comprising 10th-order Linear Prediction Coding (LPC) coefficients, frame energy, residual energy, and their respective moving averages. The GMM classifies the input into one of five noise categories: babble, car, F16, office, or white noise.
Adaptive Search Window
Based on the classified noise type, the algorithm selects an optimal search window size $D$. Stationary noises, such as vehicle or white noise, benefit from longer windows (up to 140 frames) to provide a stable estimate. In contrast, highly non-stationary noises like babble require shorter windows (as low as 55 frames) to track rapid fluctuations in the noise floor without significant lag. To ensure smooth transitions between different window sizes and prevent audio artifacts, a long-term smoothing parameter ($\gamma = 0.99$) is applied.
Experimental Results
The proposed method was evaluated using 96 sentences from male and female speakers mixed with noise from the NOISEX-92 database at SNR levels of 5, 10, and 15 dB.
Perceptual Quality Evaluation
The ITU-T P.862 PESQ scores consistently showed that the proposed adaptive method outperforms the conventional MS method. Improvements were most notable in white noise (0.03–0.05 PESQ gain) and were also present in more challenging environments like babble and office noise.
Visual Analysis of Noise Tracking
Visual inspection of the noise power estimates confirms the superiority of the adaptive approach. As shown in the figures below, the proposed algorithm tracks the actual noise power much more closely than the conventional MS method, which often exhibits a significant lag or underestimation during noise level changes.
Conclusion
By integrating environment awareness via GMM-based noise classification, we have significantly improved the robustness of the Minimum Statistics noise power estimation algorithm. The ability to adapt the search window size in real-time allows the system to maintain high accuracy across both stationary and non-stationary noise environments, leading to measurable improvements in perceived speech quality.
Summary updated