Audio-Only Speech Enhancement (ASE): Modern research in speech enhancement has been significantly advanced by the application of DNNs. CNNs, RNNs, and hybrid models, which combine Convolutional and Recurrent elements, have been instrumental in improving ASE performance.

Challenges: ASE faces several challenges, particularly when dealing with real-world environments. 

1. One of the primary difficulties is handling diverse and unpredictable noise types, including non-stationary noise (e.g., traffic or crowd noise) that varies over time, making it hard for models to consistently perform well.

2. Additionally, real-time processing demands lightweight models with low latency, which must balance computational efficiency with enhancement quality. 

3. The generalization of SE models to unseen noise conditions and different speakers is also problematic, as models trained on specific datasets may struggle with new or varied data. 

Audio-visual Speech Enhancement (AVSE) leverages both auditory and visual information, such as lip movements, to improve speech quality in noisy environments. Unlike traditional AVSE incorporates visual cues from a speaker's facial expressions and mouth movements to better distinguish speech from background noise, making it particularly effective in challenging acoustic conditions. 

Challenges: AVSE faces several challenges. 

1. Synchronizing audio and visual streams is complex, as temporal misalignment can degrade performance. 

2. Additionally, the increased computational load due to processing two data streams requires efficient, lightweight model designs, especially for real-time applications on resource-constrained devices. 

3. Another significant challenge is ensuring robustness to occlusions or poor lighting, where visual information may be compromised, requiring models to effectively prioritize audio in such scenarios.