Vocal Intensity Conversion with Continuous SPL Control via Adversarial Training
The research developed a lightweight converter that transforms speech features to control vocal intensity based on continuous decibel levels (SPL) rather than reference recordings. By training this model as a conditional GAN, it successfully renders recorded speech at a target intensity while preserving the speaker's identity and content. Experiments showed the model achieves 83% of the target intensity change, with perceived level increasing as the target level rises.