IndexTTS

Bilibili's IndexTeam builds this zero-shot speech synthesizer, which copies a speaker's timbre from one reference clip with no fine-tuning. The current IndexTTS-2.5 model, released in August 2026 on a GPT backbone of roughly 0.8 billion parameters, covers Arabic, Chinese, English, Japanese, and Spanish, keeps voice identity separate from emotion, and takes emotion cues from a second audio clip, an eight-value vector, or a written description. Pronunciation can be pinned with Pinyin, CMU phonemes, or Kana, and a duration factor stretches or compresses delivery between half and double length; IndexTTS-2 and 1.5 remain loadable. The repo ships a browser web UI, a Python API, and a vLLM recipe for production serving. Setup runs through the uv package manager with weights from Hugging Face or ModelScope, a CUDA 12.8 or newer toolkit on NVIDIA hardware, and optional DeepSpeed and half-precision modes. Licensing follows bilibili's custom Model Use License Agreement, royalty-free for commercial work, except that firms above 100 million monthly users or RMB 1 billion revenue need a separate grant. Good for dubbing and narration builders.