An Industrial-Grade Voice Cloning Sample

On August 10, Bilibili's IndexTeam released the weights for IndexTTS-2.5. It is a zero-shot TTS system: given a single reference audio clip, it clones the corresponding voice across five languages — Chinese, English, Japanese, Spanish, and Arabic — and supports cross-lingual voice transfer: feed it a Chinese reference clip, and the same voice speaks English.

Compared with the previous IndexTTS-2, this version adds Japanese, Spanish, and Arabic, infers faster, and adds speaking-speed control plus finer pronunciation control. A technical report is available as arXiv paper 2601.03888.

Architecture: A Three-Stage Pipeline

The model card discloses a classic three-stage design: a GPT backbone with roughly 0.8B parameters, a flow-matching speech-to-mel decoder, and a BigVGAN vocoder, outputting 22.05 kHz waveforms. Inference requires an NVIDIA GPU and Python 3.10-3.11, with VRAM usage around 6GB — a single consumer graphics card suffices. On first run it pulls auxiliary models including w2v-bert-2.0, MaskGCT, and CAMPPlus.

Control Is the Real Story

The selling point of IndexTTS-2.5 is controllability, across three dimensions:

  • Emotion control: emotion is disentangled from timbre and driven by an 8-float vector, fixed in the order happy, angry, sad, afraid, disgusted, melancholic, surprised, calm. To make a line like -hide quickly, he is coming- sound teary, set the sad dimension to 0.8. Text-description-based emotion requires loading an additional QwenEmotion model.
  • Pronunciation control: Chinese supports Pinyin annotation for polyphonic characters (the model card demonstrates a sentence where the same character is forced to read XING2 in one word and HANG2 in another), English uses CMU phonemes, and Japanese uses Kana readings.
  • Speaking speed: a duration_factor parameter ranging from 0.5 to 2.0, where values above 1 slow speech down and values below 1 speed it up.

An Honest Ledger on License and Ethics

A few details deserve attention. First, the license is the bilibili Model Use License Agreement — not a standard open-source license like Apache or MIT; read the terms carefully before commercial use. Second, the model card's Limitations section is blunt: the model does not verify that the speaker in a reference clip consented to being cloned — obtaining consent is the user's responsibility. At a time when voice-cloning abuse is a live controversy, that disclaimer is itself a footnote on the state of the industry. Third, long texts are segmented and concatenated with short silences, so prosody does not carry across segment boundaries; enabling random emotion sampling also reduces cloning fidelity.

Within a week of release, the model has reached 4,176 downloads in the past month on Hugging Face, with 3 community fine-tunes and 1 quantized variant already circulating.

So What

The open-source TTS track has been unusually lively lately: Qwen3-TTS, VoxCPM2, and ViiTorVoice have taken the stage in turn, and IndexTTS-2.5 offers arguably the most complete set of engineering knobs — the emotion vector plus pronunciation control. For developers, running five-language cloning in 6GB of VRAM is a low enough bar; the real question is non-technical: did the person in your reference clip say yes?

Source: Hugging Face model card (https://huggingface.co/IndexTeam/IndexTTS-2.5)