IndexTTS 2.5 Enables Cross‑Language Voice Synthesis with Zero‑Sample Multilingual Support

We introduce IndexTTS 2.5, the latest release from Bilibili’s Index Speech team, which adds zero‑sample voice cloning for five languages (Chinese, English, Japanese, Spanish, Arabic), doubles inference speed, offers controllable speech rate, and refines emotional and timbre realism compared with the previous version.

Why 2.5?

After the strong community reception of IndexTTS 2, users reported two main pain points: the need for more than just Chinese‑English support and the model’s latency in real‑time scenarios. These insights guided the development of IndexTTS 2.5.

Key upgrades

Cross‑language emotion replication with higher stability.

One‑stop support for Chinese, English, Japanese, Spanish, and Arabic.

Lightweight inference architecture that improves runtime efficiency.

Flexible speech‑rate control.

Upgrade 1 – Multilingual modeling challenge

The hardest problem in multilingual TTS is handling shared characters, e.g., Chinese characters that also appear in Japanese. To answer this, we compared three modeling strategies (details in the accompanying report) and found that no single approach is a silver bullet; each excels in different aspects and can be combined.

We built an automated data pipeline that includes event‑aware VAD segmentation, integrated ASR with speaker logs and punctuation restoration, selective source‑separation using Demucs, speaker‑consistent segment merging (single speaker, ≤25 s), and dual filtering of audio and text quality. In total we collected about 150 000 hours of multilingual audio and 135 hours of emotional speech.

Upgrade 2 & 3 – Model slimming

Two “punches” were applied:

First punch: Reduce the semantic codec frame rate from 50 Hz to 25 Hz, halving the token sequence length and cutting training/inference compute and memory with minimal loss of linguistic information.

Second punch: Replace the U‑DiT backbone in the S2M module with Zipformer, which has fewer parameters and generates faster. Subjective tests showed Zipformer preferred by 56 % vs. 40 % for U‑DiT.

On the same A10 hardware, the T2S module’s real‑time factor (RTF) dropped from 0.232 to 0.119 and the S2M module’s RTF from 0.078 to 0.017, yielding an overall speed‑up of 2.28× with no perceptible quality loss.

Upgrade 4 – GRPO post‑training

We introduced GRPO reinforcement learning into the T2S module after the main training. For each input text the model generates four candidate utterances; an ASR‑based word error rate (WER) and speaker similarity score act as judges, guiding the model toward “correct and speaker‑like” outputs.

Results: average WER across five languages fell from 6.75 % to 6.00 %**, and average speaker similarity rose from 73.18 to 73.63 , with especially notable gains for English, Japanese, and Arabic.

Performance results

Zero‑sample voice cloning: Across the five languages, IndexTTS 2.5 achieved the highest average speaker similarity, while its WER was slightly higher than the 4B Fish Audio S2 Pro model, placing it in the top tier.

Cross‑language synthesis: Using a Chinese reference utterance, the same voice could speak English, Japanese, Spanish, and Arabic, achieving the best overall scores among compared models.

Emotion transfer: Although emotion training used only Chinese and English data, the model transferred emotion to all languages; Arabic emotional tests (prompted with Japanese extremes) achieved a MOS of 4.18/5. On the CV3‑Eval zero‑shot benchmark, Chinese emotion recognition accuracy reached 0.713, far above CosyVoice3’s 0.467.

In short, a happy Chinese input yields a happy Spanish output, demonstrating clear decoupling of emotion, timbre, and language.

Conclusion

From IndexTTS 1 to 2 and now 2.5, each iteration has been driven by community feedback. IndexTTS 2.5 brings creators closer to “any voice, any language, any emotion” by delivering robust multilingual, fast, and emotionally faithful speech synthesis.