TTS-Q: Perception-Guided Post-Training Quantization of Text-to-Speech Diffusion Transformers
TTS-Q addresses how post-training quantization degrades speaker quality in large text-to-speech models by introducing a speaker-aware method. It combines calibration sampling guided by speaker embeddings with identity distillation using a full-precision teacher's speech. This approach recovers high speaker similarity and naturalness even when using low 4-bit quantization, proving it preserves voice timbre without sacrificing intelligibility.