Disentangling Pitch and Creak for Speaker Identity Preservation in Speech Synthesis

Researchers have developed a new speech synthesis system that can modify vocal 'creak'—the rough, low-pitch quality in voices—while perfectly preserving the speaker's identity. Using a conditional continuous normalizing flow technique, the model disentangles pitch from creak during training. Experiments show it significantly improves speaker verification performance across various creak manipulation strengths, achieving more natural-sounding voice modifications that don't compromise who the speaker sounds like.