PilotTTS: A Disciplined Modular Recipe for Competitive Speech Synthesis

Researchers introduced PilotTTS, an open-source autoregressive text-to-speech system trained on 200K hours of openly processed data.

It uses a multi-stage data pipeline and Q-Former-based conditioning to separate speaker identity from speaking style.

PilotTTS supports zero-shot voice cloning, emotion and paralinguistic synthesis, and Chinese dialect generation.

The model reports strong results on Seed-TTS Eval, showing competitive quality with a compact architecture.

The team released the full pipeline, pretrained weights, and code, making the recipe reusable for other groups.