PERFORMANCE EVALUATION OF ADAPTIVE FILTERS FOR SPEECH ENHANCEMENT ACROSS REALISTIC ACOUSTIC CONDITIONS

Abstract

Speech enhancement plays a critical role in improving the intelligibility and quality of speech signals in real-world acoustic environments, especially for applications such as mobile communications, hearing aids, and voice-controlled systems. This thesis presents a comprehensive study on adaptive filtering techniques for speech denoising, with a particular focus on evaluating and improving their performance in realistic noise conditions. In the first part of this work, fifteen adaptive filters from the Python Padasip toolbox are rigorously evaluated across eight real-life noise scenarios—including babble, car, exhibition hall, and airport noise—at two challenging signal-to-noise ratio (SNR) levels (5 dB and 10 dB). The performance of each filter is assessed using established objective speech quality metrics: PESQ (Perceptual Evaluation of Speech Quality), fwsegSNR (Frequency-Weighted Segmental SNR), LLR (Log-Likelihood Ratio), and COVL (Composite Objective Measure). Results reveal that while the Recursive Least Squares (RLS) filter consistently delivers superior performance, filters such as GMCC, AP, and VSLMS also demonstrate notable strength in specific noise cases or under certain evaluation criteria. This analysis provides valuable insights into the behavior of different adaptive filters and forms a benchmark for future research in the field. Building upon these findings, the second part of the thesis introduces an ensemble- based adaptive filtering approach tailored for in-car noise environments. This method dynamically combines the outputs of three filters—NLMS, GMCC, and VSLMS (Mathews’ adaptation)—using a performance-weighted scheme where filters with lower error contribute more to the final output. Additional signal processing techniques, including noise estimate subtraction, pre-emphasis, and de-emphasis, are incorporated to further suppress residual noise. Experiments conducted on in-car noisy speech samples from the NOIZEUS corpus at 5 dB and 10 dB SNR levels demonstrate that the proposed ensemble method significantly outperforms individual filters and static combinations across all objective quality measures. Together, these contributions offer a dual perspective: a detailed comparative evaluation of adaptive filters in diverse noise conditions and a novel ensemble-based enhancement system optimized for automotive noise. This work lays the groundwork for future advancements in adaptive speech enhancement systems suitable for real- time deployment in noisy environments.