Enhancing Neural Speech Embeddings for Generative Speech Models

We explore a speech enhancement framework where neural speech embeddings, derived from pre-trained self-supervised learning (SSL) models applied to noisy signals, are used as inputs to a neural vocoder to generate the corresponding clean speech. The primary innovation lies in enhancing these latent neural embeddings to mitigate distortions caused by noise and reverberation, resulting in a superior quality of the synthesized signal. By dividing the process into Separate phases for embedding enhancement and speech generation, the approach allows for greater flexibility in network design. We also examine the advantage of integrating hidden states from the SSL model in a learnable manner to create a more robust embedding for the vocoder input. Additionally, we investigate various loss functions for training the neural vocoder. Experimental results confirm the effectiveness of our proposed approach, particularly in environments with simultaneous background noise and reverberation.