hash2430/pitchtron

Implements dual prosody transfer approaches (hard and soft) that enable pitch-contour-based style transfer for Korean dialects and emotional speech without requiring parallel training data or style-specific training sets. Hard pitchtron enforces strict pitch matching when sentence structures align, while soft pitchtron pursues naturalness even with entirely different reference and target sentences through pitch-range scaling and per-phoneme control. Built on multi-speaker Korean TTS pipelines with preprocessing support for converting raw audio formats and integrating multiple Korean speech datasets.

157 stars. No commits in the last 6 months.