Beyond U-Net: A Latent-Representation-Aligned Skip-Free Backbone for Flow-Matching Speech Enhancement

The article presents a novel skip-free encoder-decoder backbone for flow-matching speech enhancement that utilizes Latent Representation Alignment (LRA) to improve the efficiency of the process. By avoiding U-Net skip connections, the model aligns its representations with clean latent features from a Descript Audio Codec, enabling compact clean-speech representation and real-time inference with only five function evaluations. Benchmark results demonstrate enhanced PESQ and perceptual quality on datasets like WSJ0-CHiME3 and VoiceBank-DEMAND, making it a significant advancement for practitioners focused on efficient speech enhancement techniques.