ByteDance's Seed-TTS is one of the strongest signals yet that TTS is graduating from a component into a genuine foundation model. This review walks through the whole family — a large-scale autoregressive stack that hits speaker similarity and naturalness scores essentially indistinguishable from ground-truth human speech in both objective and subjective evals, plus a fully diffusion-based non-autoregressive sibling called Seed-TTS_DiT that skips pre-estimated phoneme durations entirely and does speech generation through end-to-end processing. Fine-tuning pushes the subjective scores even higher, with strong controllability over emotion and other expressive attributes for speakers in the wild.

Beyond the headline quality numbers, the interesting bits are architectural and methodological: a self-distillation recipe for speech factorization that lets you cleanly disentangle attributes, and a reinforcement learning fine-tuning loop that pushes robustness, speaker similarity, and controllability further than supervised training alone. The DiT variant also unlocks speech editing in a way older NAR systems couldn't touch. If you're evaluating what a modern speech foundation model looks like — for voice cloning, expressive synthesis, dubbing, or in-context speech learning — five minutes with this review will orient you fast.