Skip to content

Seed-TTS

A research family of large autoregressive speech models

Visit official site

What Seed-TTS is

The technical report page describes Seed-TTS as a family of large autoregressive text-to-speech models from ByteDance. It reports speech that is difficult to distinguish from human speech, with strengths in in-context learning, speaker similarity and naturalness. The report also describes a self-distillation approach for speech factorisation, reinforcement learning for robustness and control, and a non-autoregressive variant, Seed-TTSDiT, built on a fully diffusion-based architecture without pre-estimated phoneme durations.

What you can do with it

  • Generate natural speech for a voice assistant
  • Produce fluent reading for audiobook work
  • Study controllable and robust speech generation

What to watch out for

  • Subjective scores are said to improve further after fine-tuning
  • The documented steps include registering for API access

Pros & cons

✓ What we like

  • The model family and its variant are described with technical specifics
  • Control over emotion and other attributes is named
  • Both autoregressive and diffusion-based directions are covered

! What to watch out for

  • No hosted endpoint, price or licence is published here
  • Performance statements come from the authors' evaluations
  • Consent handling for imitated voices is not addressed

FAQ

What is Seed-TTSDiT?

The report describes a non-autoregressive variant using a fully diffusion-based architecture.

What attributes are said to be controllable?

Emotion and other speech attributes are named in the report.

Is there a published API?

The source mentions API access in steps but publishes no endpoint or terms.

Last reviewed: 2026-11-15

More AI audio tools tools

View all →

How we review