Abstract
This paper presents a novel data-driven approach to single-channel speech enhancement employing Gaussian process (GP) regression. Our approach is based on applying GP regression to estimate the "residual gain"—the difference between the optimal gain and that obtained from the traditional minimum mean-square error log-spectral amplitude (MMSE-LSA) estimator. The proposed system involves a cascaded structure consisting of two stages. In the first stage, the gain of the MMSE-LSA estimator is calculated. In the second stage, residual gains are estimated through GP and used to further refine the output. Experimental results demonstrate that the proposed approach produces superior speech quality compared to both traditional statistical modules and other data-driven techniques.
Introduction
Speech enhancement aims to improve the quality and intelligibility of audio signals degraded by noise, such as background chatter or environmental sounds. Traditionally, this problem has been addressed using statistical models like the MMSE-LSA estimator. While effective under specific mathematical assumptions, these models often struggle with the complex, non-stationary noise found in real-world environments.
Recent shifts toward data-driven machine learning methods offer more flexibility. However, many existing techniques rely on rigid structures like codebooks or look-up tables that fail to capture the highly non-linear relationships between noise levels and optimal gain. This research bridges that gap by utilizing Gaussian Process (GP) regression, a powerful non-parametric Bayesian tool, to model these complex functions.
System Architecture and Methods
The core innovation of this work is a "cascaded" two-stage architecture that combines classic signal processing with modern regression techniques.
The Two-Stage Process
- Preliminary Stage: The system first applies the standard MMSE-LSA algorithm. This stage serves as a sophisticated feature extractor, providing an initial estimate of the speech and calculating signal-to-noise ratio (SNR) features.
- GP Regression Stage: Rather than predicting clean speech from scratch, the model focuses on the residual gain. This is the mathematical error left behind by the first stage. By learning to correct the specific errors of a known algorithm, the machine learning task becomes significantly more efficient.
Technical Innovations
To make Gaussian Processes computationally feasible for real-time speech processing, the authors implemented several optimizations:
- Clustering for Scalability: Since GP complexity scales cubically with data size, the authors used Gaussian Mixture Models (GMMs) to cluster training data into 64 groups. A smaller, faster "local expert" GP was then trained for each cluster.
- Spectro-Temporal Features: The model analyzes a window of neighboring time-frequency points. To manage this data, Principal Component Analysis (PCA) was used to compress 30 raw features into the 10 most significant components.
Experimental Setup
The system was trained using the TIMIT database (50 speakers) and tested on 10 speakers. Noise conditions were sourced from the Noisex92 database:
- Matched Condition: Training and testing both utilized White Gaussian Noise.
- Mismatched Condition: To test generalization, the model was trained on white noise but tested against F-16 cockpit noise and factory noise.
Key Results
The GP-based method consistently outperformed both the standard MMSE-LSA and baseline data-driven methods like Vector Quantization (VQ) across multiple objective metrics.
- Objective Quality: In matched white noise at 0 dB SNR, the GP method achieved a PESQ score of 0.95, significantly higher than the 0.83 achieved by MMSE-LSA.
- Robustness: In mismatched environments (e.g., F-16 noise at 0 dB), the GP method maintained its lead with a score of 2.09 versus 1.99 for the standard method.
- Visual Clarity: Spectrogram analysis revealed that the GP method effectively suppressed "residual noise"—the digital chirping artifacts often introduced by other filters—resulting in a cleaner audio profile.
Conclusion and Future Work
By focusing on error-driven learning (residual gain) and utilizing the non-linear mapping capabilities of Gaussian Processes, this approach offers a significant step forward in speech enhancement. While the current model treats frequency bins independently, future work will explore "frequency banding" to exploit the natural harmonic correlations in human speech, potentially further reducing computational load and improving audio fidelity.
Summary updated