Voice cloning with controls you can actually name
IndexTTS starts with a familiar promise: provide a short reference recording and some text, then generate speech in that voice. Version 2.5 adds five-language output and direct controls for emotion, timing, and pronunciation. It supports Chinese, English, Japanese, Spanish, and Arabic, including cross-language cloning.
Emotion is not limited to a vague style prompt. You can supply a second recording as the emotional reference, adjust its influence, provide an eight-value vector covering moods such as happy, angry, sad, surprised, or calm, or ask a Qwen-based component to infer emotion from separate text. The project says random sampling can reduce cloning fidelity, a useful warning for anyone balancing expression against identity. Version 2.5 also exposes a duration factor from half to twice normal duration.
Pronunciation markup addresses a problem that matters in real scripts. Chinese text can disambiguate characters with Pinyin, English can insert CMU phonemes, and Japanese can specify Kana readings. This beats respelling a name and hoping the model cooperates. The current scheme is not universal: the README says older Pinyin control only supports valid combinations from its vocabulary, and an open issue says English contractions such as we'll are still read incorrectly. Build a pronunciation dictionary around your actual copy.
Installation is reproducible, not small
The project has moved to uv and asks users to let it manage Python and dependencies in a project environment. The main sequence is sensible: clone, run uv sync --all-extras, download IndexTTS-2.5 weights from Hugging Face or ModelScope, check GPU detection, then start the WebUI. Examples and several auxiliary models download on demand. A current pull request notes that these extra caches include Wav2Vec-BERT, a codec, CAMPPlus, and BigVGAN, and that the main model download is not a complete installation.
Current package metadata supports Python 3.10 and 3.11. A pull request demonstrates newer Python working with dependency updates, but that is not the released constraint. Linux and Windows users are told to install CUDA 12.8 or newer if installation reports CUDA trouble. DeepSpeed is optional, difficult to install on Windows, and may make inference slower depending on the machine. Half precision reduces memory with a stated small quality cost, but hardware still determines whether this feels interactive.
The Gradio interface is the easiest evaluation route. The Python examples show every major conditioning mode. Production deployment points to an official vLLM recipe rather than treating the demo UI as a service. Teams should confirm which 2.5 functions the serving path exposes before designing an API around WebUI examples.
Version 2.5 is powerful and freshly integrated
The August 10, 2026 update changes more than the model file. It adds languages, speed control, newer pronunciation syntax, and different BF16 behavior. The repository was pushed on August 12, and current pull requests are actively correcting version routing, installation choices, cache locations, and output handling. The newest path has not settled.
One open report says the published 2.5 configuration does not match the expected setup and the WebUI fails with missing models. A proposed CLI fix says the existing indextts entry point always constructs the version 1 inferencer even when a version 2 configuration is supplied. Another proposed fix describes severe clipping when generated audio is passed to newer TorchAudio save behavior with the wrong numeric scaling. The pinned TorchAudio 2.8 path avoids a related 2.9 compatibility report, but downstream packagers can easily step outside that safe combination.
Test the exact route you will ship. Generate through the API, CLI, and server separately if you expose all three. Check saved files for peak saturation, silence, duration, and truncation rather than trusting a successful return code. Keep known sentences with contractions, names, numbers, mixed languages, and each required emotion.
The model license changes the buying decision
GitHub cannot map the license to a standard SPDX identifier because IndexTTS uses the bilibili Model Use License Agreement. It grants wide use but requires a separate license when a user or affiliate exceeds 100 million monthly active users or RMB 1 billion in prior-year revenue. It also places obligations on derivatives and downstream recipients, limits use of the model to improve other commercial AI models, and selects Chinese law and Shanghai arbitration. The Chinese license text prevails if translations conflict.
The separate disclaimer is stricter in practical voice use. It prohibits synthesizing politicians, public figures, or any unauthorized person, as well as fraud, false information, and unauthorized commercial use of generated voices. A responsible product needs documented speaker consent, access controls, output labeling where required, and a process for removal requests. Do not treat a user-uploaded WAV as proof of authorization.
There is no official training workflow in the main documentation. An issue links an unofficial version 2 training branch for multilingual adaptation, but its author says it has no instructions and may remain undocumented. Teams that need supported fine-tuning should compare CosyVoice or a broader speech toolkit.
Excellent evaluator, cautious production pick
The latest GitHub release object is still version 1.5 from September 2025, but that tag is poor evidence of current health. The repository announces 2.5, links current model files and papers, and had code and issue activity on August 12, 2026. Its 381 open items include issues and pull requests, so the count reflects a busy queue rather than 381 confirmed defects.
IndexTTS-2.5 deserves a trial for multilingual cloning that needs directed emotion and pronunciation. Its WebUI makes comparison approachable and its Python examples expose the useful details. For production, pin dependencies, audit the custom license, collect consent, and listen to regression samples on every update. The model is promising; the surrounding 2.5 software still demands engineering supervision.