Enhancing Speech Quality with Wave-U-Net and GANs

Recent advancements in deep learning have significantly improved speech enhancement techniques, crucial for various audio applications. This paper presents a novel method that integrates the Wave-U-Net model with Generative Adversarial Networks (GANs) to enhance speech quality. The Wave-U-Net architecture, known for its efficiency in processing audio signals via convolutional layers, is combined with GANs to create a more robust speech enhancement model. This approach aims to effectively reduce noise while improving speech’s intelligibility and perceptual quality. To train the model, noisy speech samples with varying SNRs were generated using the LibriSpeech dataset for clean speech and the NOISEX-92 dataset for synthetic noise. The performance of the proposed model was evaluated based on SNR improvements, with results showing enhancements of up to 6.7 dB.