A Comparative Evaluation of Deep Learning Models for Speech Enhancement in Real-World Noisy Environments

Speech enhancement, particularly denoising, is vital in improving the intelligibility and quality of speech signals for real-world applications, especially in noisy environments. While prior research has introduced various deep learning models for this purpose, many struggle to balance noise suppression, perceptual quality, and speaker-specific feature preservation, leaving a critical research gap in their comparative performance evaluation. This study benchmarks three state-of-the-art models Wave-U-Net, CMGAN, and U-Net, on diverse datasets such as SpEAR, VPQAD, and Clarkson datasets. These models were chosen due to their relevance in the literature and code accessibility. The evaluation reveals that U-Net achieves high noise suppression with SNR improvements of +71.96% on SpEAR, +64.83% on VPQAD, and +364.2% on the Clarkson dataset. CMGAN outperforms in perceptual quality, attaining the highest PESQ scores of 4.04 on SpEAR and 1.46 on VPQAD, making it well-suited for applications prioritizing natural and intelligible speech. Wave-U-Net balances these attributes with improvements in speaker-specific feature retention, evidenced by VeriSpeak score gains of +10.84% on SpEAR and +27.38% on VPQAD. This research indicates how advanced methods can optimize trade-offs between noise suppression, perceptual quality, and speaker recognition. The findings may contribute to advancing voice biometrics, forensic audio analysis, telecommunication, and speaker verification in challenging acoustic conditions.

💡 Research Summary

This paper presents a systematic comparative study of three state‑of‑the‑art deep learning models for speech enhancement—Wave‑U‑Net, Conditional GAN (CMGAN), and U‑Net—under realistic noisy conditions. The authors identify a gap in the literature: while many recent works propose powerful architectures, they often prioritize a single aspect such as noise suppression, perceptual quality, or speaker‑specific feature preservation, leaving the trade‑off between these criteria insufficiently explored. To address this, the study benchmarks the three models on three publicly available, acoustically diverse evaluation sets: SpEAR (synthetically mixed noise), VPQAD (real‑world adult speech with environmental noise), and the Clarkson dataset (real‑world child speech recorded in outdoor environments).

Model architectures are described in detail. Wave‑U‑Net adopts a 1‑D convolutional U‑shaped network that encodes the raw waveform at multiple temporal scales and reconstructs it via skip connections, allowing global context capture while preserving fine‑grained details. CMGAN is a conditional generative adversarial network that receives explicit noise‑type conditioning; the generator learns to map noisy inputs to clean speech, while the discriminator enforces realism, resulting in outputs that sound natural to human listeners. U‑Net follows the classic encoder‑decoder design borrowed from image segmentation, employing multi‑scale feature fusion to separate speech from noise effectively.

Training data are assembled from five large corpora: DEMAND (a wide variety of environmental noises), MUSDB18‑HQ (music‑speech separation), VCTK (multiple speakers), LibriSpeech (large read speech corpus), and ESC‑50 (environmental sound classification). This combination provides a rich mixture of noise types, speaker variability, and recording conditions, ensuring that the models learn robust representations applicable to real‑world scenarios.

Evaluation employs three complementary metrics. Signal‑to‑Noise Ratio (SNR) improvement quantifies objective noise reduction. Perceptual Evaluation of Speech Quality (PESQ) estimates human‑perceived audio quality. VeriSpeak, a speaker verification score, measures how well speaker‑specific characteristics are retained after enhancement. By reporting all three, the authors capture a holistic view of performance.

Results reveal distinct strengths for each model. U‑Net achieves the highest SNR gains: +71.96 % on SpEAR, +64.83 % on VPQAD, and a remarkable +364.2 % on the Clarkson dataset, indicating superior noise suppression across both synthetic and real recordings. CMGAN attains the best PESQ scores, 4.04 on SpEAR and 1.46 on VPQAD, demonstrating that adversarial training preserves naturalness and intelligibility. Wave‑U‑Net shows the largest improvements in VeriSpeak, +10.84 % on SpEAR and +27.38 % on VPQAD, suggesting that its skip‑connection design maintains fine‑grained speaker cues essential for biometric applications.

These findings underscore the inherent trade‑offs among noise reduction, perceptual quality, and speaker identity preservation. For latency‑critical communication systems where maximal denoising is required, U‑Net appears most suitable. In media streaming or teleconferencing where listener experience is paramount, CMGAN offers the best perceptual outcome. For security‑sensitive domains such as voice biometrics, forensic audio analysis, or speaker verification under adverse conditions, Wave‑U‑Net provides the most reliable retention of speaker‑specific features.

The paper also discusses practical implications. By using publicly accessible code and datasets, the study promotes reproducibility and facilitates further research. The authors suggest future directions including model compression for real‑time deployment, multi‑task learning to jointly optimize SNR and PESQ, and expanding training data to cover more languages, dialects, microphone types, and moving‑speaker scenarios.

In summary, this work delivers a comprehensive benchmark that clarifies how advanced deep learning models can be selected and tuned according to specific application priorities in noisy acoustic environments, thereby advancing the state of speech enhancement technology for real‑world use cases.