IndexTTS
Bilibili's industrial-grade zero-shot TTS: voice cloning with pinyin-level Chinese pronunciation control.
About
IndexTTS is a GPT-style zero-shot text-to-speech system from Bilibili's IndexTeam, with weights and inference code released March 25, 2025 and a paper on arXiv (2502.05512). It builds on XTTS and Tortoise with a conformer-based speaker encoder and BigVGAN2, trained on tens of thousands of hours of speech. Its differentiator is control: Chinese polyphonic characters can be corrected with pinyin, pauses can be placed anywhere with punctuation, and later versions added emotion control and precise duration control.
The team iterates fast: 1.5 landed May 2025 with better English, 2.0 in September 2025 with duration control, and 2.5 in August 2026 with five languages and vLLM serving. The provider claims state-of-the-art results over XTTS, CosyVoice2, Fish-Speech, and F5-TTS, but those are its own comparisons. Note the licensing: weights and code are public, but commercial use requires contacting Bilibili, so it is not a standard OSI-approved open license.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
No independent public benchmark table found; provider claims state-of-the-art over XTTS, CosyVoice2, Fish-Speech, and F5-TTS (provider-reported, arXiv:2502.05512)
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (metrics tier, metrics-v1). Evidence: real Bilibili IndexTeam release Mar 25, 2025 with public weights, code, paper (arXiv:2502.05512), and demos; rapid iteration through 2.5 (Aug 2026) shows active maintenance. Open weights. Release facts corroborated across the paper, HF, and community mirrors. Strengths: strong controllability story (pinyin, duration, emotion), active development. Docked for: (1) quality claims are provider-reported with no independent benchmark found; (2) commercial use requires a Bilibili license, so it is not standard open source.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Openness · license + docs
- 6.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · Bilibili IndexTeam; weights+code Mar 25 2025; paper arXiv:2502.05512
- · Rapid iteration: 1.5 (May 2025), 2.0 (Sep 2025), 2.5 (Aug 2026, vLLM serving)
- · Strong controllability: pinyin, duration, emotion control
- · Quality claims are provider-reported; commercial use needs Bilibili license (not OSI open)