What is IndexTTS-2.5?
IndexTTS-2.5 is a zero-shot text-to-speech model from IndexTeam, the speech group at the Chinese video platform Bilibili, with weights posted in August 2026. It clones a voice from a single reference clip and speaks Chinese, English, Japanese, Spanish and Arabic, carrying the voice across languages. A GPT-style backbone of about 0.8 billion parameters predicts speech tokens, a flow-matching decoder turns them into a spectrogram and a BigVGAN vocoder outputs 22.05 kHz audio.
Its distinctive feature is that emotion is controlled separately from the voice: you can pass an eight-value vector (happy, angry, sad, afraid, disgusted, melancholic, surprised, calm), a second audio clip, or a text description. Over IndexTTS-2 it adds three languages, speaking-speed control and readings written inline in Pinyin, CMU phonemes or Kana, and the technical report claims a 2.28-times faster real-time factor.
Is IndexTTS-2.5 worth it?
IndexTTS-2.5 is worth its setup for one reason: it lets you change how a cloned voice feels without changing who it sounds like. Most open cloners copy the emotion of the reference clip along with the timbre; here an eight-value vector, a second clip or a sentence sets the emotion independently, and speed can be stretched from 0.5 to 2 times.
What it gains in control it gives up on breadth and on independent proof. It covers five languages, against 30 for VoxCPM2 and over 100 for Higgs TTS 3. The technical report claims error rates and speaker similarity comparable to IndexTTS-2 but names no benchmark scores in its abstract, and no independent arena lists it. Rivals' tables give a hint about the family: Boson's EmergentTTS-Eval puts IndexTTS-2 at 31.12% against 53.65% for Higgs TTS 3, and its cross-lingual error rates on CV3 ran above 100%.
It needs a dedicated GPU but not a large one. Bilibili asks for an NVIDIA GPU, Python 3.10 or 3.11 and roughly 6 GB of VRAM; auxiliary models including w2v-BERT 2.0 and the MaskGCT codec download on first run. Text-described emotion requires loading a separate Qwen-based emotion model.
The catch: the licence is permissive in price but not in scope. It is free for commercial use below 100 million monthly active users and RMB 1 billion in revenue, yet it bans using the model to improve other AI models, falls under PRC law with arbitration in Shanghai, and lets Bilibili revoke it on breach. The model card says plainly that it does not check whether the speaker in a reference clip consented. Long text is cut into segments joined by short silences, so prosody does not flow across paragraphs.
Our take: for anime-style dubbing or games in Chinese, Japanese or English, IndexTTS-2.5's emotion dials are the main event, and a lawyer should read the licence before anyone ships with it.
Where it earns its place
A dubbing studio localising Japanese anime into Chinese
Keep each character's voice while matching scene-by-scene emotion
Clone each actor from one clip, then pass the original line's audio as the emotion reference, so a shout stays a shout in the new language. Runs on a single GPU with about 6 GB of VRAM.
An indie game developer writing branching dialogue
Render the same line as calm, angry or afraid without re-recording
Set the eight-value emotion vector per line, for example 0.8 on sad, and batch-render variants. Free under the licence far below the 100 million-user threshold.
A language-learning app producing Japanese and Spanish drills
Control exact readings of tricky words
Write readings inline in the <word|reading> form with Kana or CMU phonemes, and slow speech to 1.2 to 1.5 with the duration factor for beginners.
Before you commit
- Setup
- Clone index-tts/index-tts, run uv sync, download weights from Hugging Face or ModelScope; Python 3.10 to 3.11, NVIDIA GPU, roughly 6 GB of VRAM. A Gradio web UI is included.
- Integrations
- Python API and web UI only; no first-party hosted API or OpenAI-style server is documented.
- Data and privacy
- Self-hosted, so clips and text stay local. The bilibili licence makes you responsible for personality and data-protection rights; the model card says it does not verify consent for cloned voices.
- Cost at scale
- No fee below 100 million monthly active users or RMB 1 billion revenue; above that Bilibili must grant a separate licence at its discretion. Otherwise the cost is GPU time.
- Limits
- Five languages; 22.05 kHz output; long text is split into segments with no prosody across boundaries; random emotion sampling lowers cloning fidelity.
What’s good
- Emotion controlled separately from voice, by vector, clip or text
- Free for commercial use below 100 million monthly users
- Runs in roughly 6 GB of VRAM
- Inline Pinyin, CMU phoneme and Kana readings fix mispronunciations
What’s not
- Five languages only
- No independent benchmark or arena placement
- Licence bans improving other models and sits under PRC law
- No consent check; long text loses prosody at segment joins
Best for
Dubbing and game-audio teams working in Chinese, Japanese or English who need to direct a cloned voice's emotion line by line.
IndexTTS-2.5 alternatives
IndexTTS-2.5: common questions
- Is IndexTTS-2.5 worth it?
- IndexTTS-2.5 is the open cloner for dubbing work where an actor's emotion has to change but their voice must not. We rate it 3.8 out of 5.
- How much does IndexTTS-2.5 cost?
- Free under the bilibili Model Use License, including commercial use, unless you have over 100 million monthly users or RMB 1 billion annual revenue. No hosted API.
- What are the downsides of IndexTTS-2.5?
- Five languages only; no independent benchmark or arena placement; licence bans improving other models and sits under PRC law; no consent check; long text loses prosody at segment joins.
- Who is IndexTTS-2.5 best for?
- Dubbing and game-audio teams working in Chinese, Japanese or English who need to direct a cloned voice's emotion line by line.
- How much does IndexTTS-2.5 cost for a team or in production?
- No fee below 100 million monthly active users or RMB 1 billion revenue; above that Bilibili must grant a separate licence at its discretion. Otherwise the cost is GPU time.
- Is IndexTTS-2.5 safe for company data?
- Self-hosted, so clips and text stay local. The bilibili licence makes you responsible for personality and data-protection rights; the model card says it does not verify consent for cloned voices.
- What are the best IndexTTS-2.5 alternatives?
- The closest alternatives we have reviewed are Fun-CosyVoice3, VoxCPM2, Qwen3-TTS and Fish Audio Drama 3.
Answers checked Nov 7, 2026.