IndexTTS-2.5 Operational

IndexTTS-2.5 is an industrial-grade zero-shot voice cloning model open-sourced by Bilibili. With only 0.8B parameters, it supports cross-lingual transfer across Chinese, English, Japanese, Spanish, and Arabic. The model restructures the inference architecture, increasing inference speed by 2.28 times. It supports speech rate adjustment from 0.5x to 2.0x, as well as fine-grained control over pronunciation, emotion, and other characteristics. The complete code, model weights, and paper have all been released under the Bilibili Model License.

Zero-shot voice cloning: With only 3 to 10 seconds of reference audio, it can accurately reproduce the timbre characteristics of any speaker without additional training for the target speaker.
Five-language cross-lingual support: Supports Chinese, English, Japanese, Spanish, and Arabic.
Fine-grained emotion control: Provides multiple ways to adjust emotion: conveying emotion through an independent emotional reference audio, precisely specifying emotional intensity with an eight-dimensional emotion vector, enabling text-driven automatic emotion inference, and adjusting emotional intensity with the emo_alpha parameter.
Disentangled timbre and emotion: The timbre prompt audio and emotion prompt audio can be specified completely independently. Users can separately control “who is speaking” and “how they are speaking.”
Continuous speech rate adjustment: The duration_factor parameter allows the synthesis speed to be adjusted freely between 0.5x and 2.0x to meet the pacing requirements of different scenarios.
Precise pronunciation intervention: Supports pronunciation control at three levels—Chinese Pinyin, English CMU phonemes, and Japanese kana—effectively addressing the pronunciation of polyphonic characters, variant pronunciations, and context-dependent pronunciation.

Follow on WeChat and reply “开源” to join the AI open-source project community
Three-stage generation pipeline: The model uses a three-stage architecture: “text → semantic Token → mel-spectrogram → waveform.” A Transformer-based text-to-semantic (T2S) language model generates semantic Tokens, a non-autoregressive semantic-to-mel-spectrogram (S2M) flow-matching module generates the mel-spectrogram, and a neural vocoder finally reconstructs the resulting audio.
Semantic codec compression: The frame rate of semantic Tokens is compressed from 50Hz to 25Hz, directly halving the sequence length and significantly reducing the computational cost and memory usage of the T2S module during training and inference. This is one of the key factors behind the inference acceleration.
Zipformer architecture upgrade: The backbone network of the S2M module has been replaced from U-DiT with the more efficient Zipformer architecture. With fewer parameters, Zipformer provides stronger long-range dependency modeling, making mel-spectrogram generation faster and more stable while preserving synthesis quality.
Cross-lingual modeling strategy: To address confusion caused by overlapping characters in multilingual scenarios, the team proposed three strategies: boundary-aware alignment (inserting language markers such as before and after the text), Token-level concatenation (fusing language-specific embeddings for each input Token), and instruction-guided generation (adding a natural-language instruction prefix such as “Please read aloud in English” before the text).
GRPO reinforcement learning optimization: During the post-training stage of the T2S module, GRPO (Group Relative Policy Optimization) is introduced. The word error rate (WER) of a frozen ASR model serves as the reward signal. Four candidate samples are generated each time, and the relative likelihood of high-reward samples is optimized to continuously improve pronunciation accuracy and speech naturalness.

Environment preparation: Install Git and the uv package manager first, then download the official code repository locally by running git clone https://github.com/index-tts/index-tts.git.
Dependency installation: After entering the project directory, run uv sync --all-extras to automatically create a virtual environment and install all required Python dependencies.
Model download: Use huggingface-cli or modelscope to download the model weights from the official repository IndexTeam/IndexTTS-2.5 into the local checkpoints directory.
Launch the Web interface: Run uv run webui.py to start the Gradio interactive interface, then visit http://127.0.0.1:7860 in a browser for visual speech synthesis operations.
Load the model: In a Python script, import the IndexTTS2 class from indextts.infer_v2_5 and pass the configuration file path and model directory to instantiate the model.
Basic voice cloning: Call the instance’s infer method, passing the reference audio path, target text, and language code to generate and save a speech file with the cloned timbre.
Emotion control: Pass the emo_audio_prompt parameter to the infer method to specify emotional reference audio, or use emo_vector to provide an eight-dimensional emotion vector controlling the emotional expression of the synthesized speech.
Speech rate adjustment: Set the duration_factor parameter to a value between 0.5 and 2.0 during inference to proportionally speed up or slow down the synthesized speech.
Pronunciation intervention: Insert annotations such as <行|XING2>, or <上手|じょうず> into the input text to precisely control the pronunciation of polyphonic characters or variant pronunciations.

High efficiency with a small parameter count: With only 0.8B parameters, the model achieves 2.28x faster inference through semantic codec compression and the Zipformer architecture upgrade. Its real-time factor (RTF) is as low as 0.20 at BF16 precision, balancing lightweight design and high performance.
Five-language cross-lingual transfer: Supports Chinese, English, Japanese, Spanish, and Arabic. Users can use reference audio in any one of these languages to clone speech in another, and achieve cross-lingual emotion transfer without emotional training data in the target language.
Fine-grained multidimensional control: The model supports continuous speech rate adjustment from 0.5x to 2.0x, eight-dimensional emotion vector control, text-driven emotion inference, and pronunciation intervention at three levels—Pinyin, CMU phonemes, and Japanese kana—to meet the fine-grained speech synthesis requirements of content creation.
Complete disentanglement of timbre and emotion: Users can independently specify the timbre prompt audio and emotion prompt audio. The two can even come from different languages, allowing the desired emotional style to be flexibly injected while retaining the target speaker’s timbre.
Open-source Arabic support: Among mainstream open-source TTS models, IndexTTS-2.5 is one of the few to fully support Arabic voice cloning and cross-lingual transfer, filling a gap for Arabic in the open-source speech synthesis field.

GitHub repository: https://github.com/index-tts/index-tts
arXiv technical paper: https://arxiv.org/pdf/2601.03888

AI dubbing and audiobook content generation: Content creators only need to provide a reference audio clip of a presenter or character to quickly generate audiobooks, radio dramas, and podcast content in multiple languages and emotional styles, without repeated recording.
Cross-language video localization: Film, television, and short-video teams can use Chinese reference audio to directly generate dubbing in English, Japanese, Spanish, or Arabic while retaining the original actor’s timbre, substantially reducing the cost of overseas distribution.
Game and virtual character voices: Game developers can clone a specific timbre for NPCs or virtual idols and use emotion vectors to switch between happiness, anger, sadness, and joy in real time, enabling dynamic narrative voices and livestream interaction.
Real-time voice interaction and digital humans: With an ultra-low real-time factor of 0.20, IndexTTS-2.5 can be deployed in smart customer service, digital-human livestreaming, and real-time translation scenarios to support low-latency personalized voice responses.
Education and language learning: Educational institutions can use pronunciation control to generate demonstration audio for standard Pinyin, CMU phonemes, or Japanese kana, and clone a teacher’s timbre to create synchronized multilingual courses.