CosyVoice vs IndexTTS vs VoxCPM: licence, cloning and cost

I have watched this go wrong more than once in Shenzhen. A team picks a Chinese open TTS model on the strength of a demo clip, wires it into a product, then finds out at legal review that the weights were never under the licence the README appeared to promise. All three of these models are good. Only two of them can be shipped without a phone call.

CosyVoice is Alibaba's FunAudioLLM line. IndexTTS comes from bilibili. VoxCPM comes from OpenBMB, the group behind the MiniCPM models. English-language coverage of all three is thin, and where it exists it tends to copy a licence summary from someone else's blog post rather than open the file. I read the Chinese. In one case the Chinese text says something materially different from the English, and the licence itself says the Chinese wins.

The licence, and the number the English text gets wrong

Two of the three are simple. The CosyVoice repository carries the unmodified Apache License 2.0 text with no appended use restrictions, and the Fun-CosyVoice3-0.5B-2512 model card carries the apache-2.0 tag on the weights. VoxCPM is the same picture stated more bluntly: OpenBMB puts "released under the Apache-2.0 license, free for commercial use" on the model card and repeats it in the README. Both cards carry ethical language about impersonation and about demo content being illustrative. That language is not in either licence file, and the licence file is the operative grant.

IndexTTS is the one worth reading properly. In July 2025 someone opened issue #228 on the repository pointing out that the root LICENSE said Apache 2.0 while a separate INDEX_MODEL_LICENSE demanded prior written authorization for commercial use. No maintainer answered. That ambiguity is what most English write-ups froze in amber, and it is why you still see "IndexTTS requires written bilibili authorization for any commercial use" repeated as fact.

As of 1 September 2026 the repository resolves it differently. There is one LICENSE, the bilibili Model Use License Agreement, alongside LICENSE_ZH.txt and a DISCLAIMER. No Apache text remains anywhere in the tree. Section 2.1 grants a worldwide, non-exclusive, non-transferable, royalty-free limited licence to use the model. There is no blanket commercial bar. Section 2.2 instead sets a scale threshold in the Meta style: you must request a separate licence only if you or an affiliate had more than 100 million monthly active users in the preceding calendar month, or if annual revenue in the preceding year crossed a stated figure.

That stated figure is where the English and Chinese diverge. The English LICENSE says RMB 1 billion. LICENSE_ZH.txt says 年收入超过1亿人民币, which is RMB 100 million, one tenth as much. Section 9 of the English text settles which one applies: where the two versions conflict, the Chinese version prevails and governs the rights and obligations of the parties. So the real trigger is RMB 100 million, and a mid-sized Chinese media company that reads only the English text will conclude it is clear when it is not. Section 6 sends any dispute to the Shanghai Arbitration Commission under PRC law.

Two more clauses matter in practice. Section 4 bars using the model or a derivative work to improve any other AI model, with a carve-out for IndexTTS itself, its derivatives and non-commercial models, so distilling it into your own commercial voice model is out. And the DISCLAIMER's line about 未经授权将合成声音用于商业目的 is about cloning a real person's voice without that person's authorization, not about commercial use of the model. English summaries routinely conflate the two. One loose end: the licence defines the model by name as "bilibili indextts2", while the current release is IndexTTS-2.5. The 2.5 model card points at the same agreement, so the intent is clear enough, but the definition has not been updated to match.

What the papers actually report

The CosyVoice 3 paper describes a 1.5B model and reports 0.71% CER with 0.836 speaker similarity on SEED test-zh, and 1.45% WER with 0.784 on test-en, after reinforcement learning. Those weights were not released. What you can download is Fun-CosyVoice3-0.5B-2512, at 0.5B, and its card publishes its own numbers: 1.21% CER and 78.0 similarity on test-zh for the base checkpoint, 2.24% WER and 71.8 on test-en, with the RL variant at 0.81% and 1.68% respectively. Quoting the paper's figures for the checkpoint you actually ran is a common and consequential error.

VoxCPM2's README reports 1.84% WER at 75.3 similarity on seed-tts-eval test-EN, and 0.97% CER at 79.5 on test-ZH. Its comparison table places IndexTTS2 at 2.23% and 1.03%. Those IndexTTS numbers were produced by the VoxCPM team, not published by bilibili, and a vendor's table of its competitors deserves the scepticism you would apply to any vendor's table of its competitors.

IndexTTS's own reporting is where comparisons break down entirely. The IndexTTS2 paper claims to beat state of the art on word error rate, speaker similarity and emotional fidelity, without numbers in the abstract. The IndexTTS-2.5 README does not report seed-tts-eval at all. It reports CV3-Eval, Alibaba's benchmark, at 6.75 WER and 73.18 speaker similarity averaged across Chinese, English, Japanese, Spanish and Arabic, improving to 6.00 and 73.63 for the RL variant. A five-language average on a harder benchmark is not the same measurement as 1.84 on English seed-tts-eval. Anyone who puts those two figures in one table is misleading you, whether or not they mean to.

The accuracy story is also not one-sided. On seed-tts-eval's hard set, CosyVoice's RL checkpoint reports 5.44% CER against VoxCPM2's 8.13%, and VoxCPM's own table gives IndexTTS2 7.12%. And the CosyVoice card publishes a human baseline row of 1.26% CER on test-zh, which the RL checkpoint beats. That tells you the metric has saturated, not that the synthesis is better than a person. Where IndexTTS genuinely leads is control: an eight-float emotion vector disentangled from timbre, and a mode that fixes the generated token count so you can hit a target duration exactly. For dubbing to picture, that is the feature that matters, and neither of the other two offers its equivalent. The card is honest that enabling random emotion sampling degrades cloning fidelity.

Languages, dialects and the accent nobody trained on

CosyVoice covers nine languages: Chinese, English, Japanese, Korean, German, Spanish, French, Italian and Russian, plus what the card calls 18 or more Chinese dialects, with Guangdong, Minnan, Sichuan, Dongbei, Shanghai, Tianjin, Shandong, Ningxia and Gansu named explicitly. It was trained on a million hours. Nothing else in the open field comes close on Chinese regional speech, and CV3-Eval's only accent subset is Chinese accent voice cloning.

IndexTTS-2.5 covers five: Chinese, English, Japanese, Spanish and Arabic, with the last three added in the 2.5 release. VoxCPM2 covers thirty, including Arabic, Hindi, Swahili, Tagalog, Vietnamese, Thai and Khmer, plus nine Chinese dialects. If your requirement is breadth, that is not a close contest.

On African-accented English, the honest answer is that none of the three documents it. CosyVoice 2's data description listed Chinese, Indian and Russian English accents as speaking styles. The CosyVoice 3 paper does not mention African accents anywhere. VoxCPM2 supports Swahili as a language, which is a different thing from Nigerian or Ghanaian English. IndexTTS says nothing on the subject.

That leaves zero-shot cloning from your own reference clip as the only route, and here is the part that should worry you: none of these projects publishes an accent-preservation metric. Speaker similarity on seed-tts-eval measures timbre against a mostly Chinese and American reference distribution. It will not tell you whether a synthesised Lagos or Nairobi voice keeps its vowel system and its even-paced rhythm, or drifts toward General American on longer passages. I have not run a controlled study, so I will not pretend to a number. What I will say is that this is the one axis you must evaluate yourself, with your own speakers and your own listeners, before you commit. Treat every published similarity figure as silent on it.

Hardware and what a second of audio costs you

All three are self-hosted, so the cost line is GPU-hours rather than per-character API billing. That changes the calculation entirely if your volume is high and your latency budget is loose.

VoxCPM2 is the heaviest at 2B parameters, roughly 8GB of VRAM and a real-time factor of about 0.30 on an RTX 4090, dropping to about 0.13 with Nano-vLLM acceleration. It needs Python 3.10 or later below 3.13, PyTorch 2.5 or later and CUDA 12. In exchange you get native 48kHz output that you do not have to upsample, plus voice design from a written description with no reference clip at all.

IndexTTS-2.5 is 0.8B and reports 0.2065 RTF at bf16 on the same RTX 4090, so roughly five times faster than real time. VRAM is not documented. CosyVoice's released checkpoint is the smallest at 0.5B and the card claims bi-streaming latency as low as 150ms, but names no reference hardware for that figure and states no VRAM requirement, so treat it as a ceiling rather than a spec. For a streaming voice agent, the streaming architecture matters more than the raw RTF, and CosyVoice is the only one of the three built around it.

When none of the three is the answer

If you are building a conversational agent that has to answer inside 200 milliseconds end to end, self-hosting any of these and then attaching ASR, an LLM and turn detection is a larger engineering programme than buying a hosted low-latency voice API. The model is the easy part of that stack.

If you need indemnification against a voice-likeness claim, none of these three gives it to you. Apache 2.0 disclaims warranties, and the bilibili agreement puts the burden of defending third-party infringement claims squarely on you. Any commercial deployment that clones identifiable people needs signed releases regardless of which model produced the audio.

And if your actual target is a low-resource African language rather than accented English, none of the three trained on it in any meaningful volume. The real project there is a fine-tune on your own corpus, which makes the licence question decisive again for a different reason: you can fine-tune and ship either Apache 2.0 model freely, while section 4 of the bilibili agreement restricts using IndexTTS or its derivatives to improve other models. That clause alone rules IndexTTS out of most serious localisation work.

Questions

Sources

1. CosyVoice repository LICENSE (Apache 2.0) github.com
2. Fun-CosyVoice3-0.5B-2512 model card, evaluation table and dialect list huggingface.co
3. CosyVoice 3: Towards In-the-wild Speech Generation via Scaling-up and Post-training (arXiv 2505.17589) arxiv.org
4. bilibili Model Use License Agreement, English text github.com
5. bilibili Model Use License Agreement, Chinese text (LICENSE_ZH.txt) github.com
6. IndexTTS repository README, benchmarks and RTF github.com
7. IndexTTS licence ambiguity, issue #228 github.com
8. IndexTTS-2.5 model card huggingface.co
9. IndexTTS2 paper (arXiv 2506.21619) arxiv.org
10. VoxCPM repository README, benchmark table and language list github.com
11. VoxCPM2 model card huggingface.co
12. VoxCPM: Tokenizer-Free TTS for Context-Aware Speech Generation (arXiv 2509.24650) arxiv.org