The recent developments in the field of deep learning-based compression and speech enhancement technologies have shown significant advancements, focusing on improving efficiency, quality, and adaptability. In video compression, there's a notable shift towards leveraging neural networks for rate control and implicit neural representations for video encoding, aiming for higher quality at lower bit rates and additional functionalities like upsampling and stabilization. Speech enhancement research is moving towards decoupling amplitude and phase information to avoid compensation effects and reduce model complexity, while also exploring novel quantization methods based on structural entropy for better speech representation in language models. These advancements indicate a broader trend of integrating deep learning techniques to overcome the limitations of traditional methods, with a particular emphasis on optimizing performance and reducing computational complexity.

Noteworthy papers include:

• A neural network-based rate control scheme for deep video compression that accurately determines coding parameters without pre-encoding, significantly improving rate control accuracy and mitigating quality fluctuations.
• A dual-path network for speech enhancement that decouples amplitude and phase information, reducing computational complexity while maintaining superior performance.
• A novel speech representation codec based on structural entropy, offering improved speech reconstruction and surpassing existing models in zero-shot text-to-speech tasks.
• A pixel-wise implicit neural representation for video compression that provides state-of-the-art results and additional video processing capabilities.
• An approach to optimize neural codecs by exploiting vector quantization and the entropy gradient, demonstrating significant rate savings and performance improvements.