Kokoro TTS alternatives: Chatterbox, Qwen3-TTS, VoxCPM

As of 2026-10-14, the cited sources document three alternatives to Kokoro TTS: Chatterbox, Qwen3-TTS and VoxCPM. Chatterbox has the most detailed documentation, including 23 languages and MIT licensing. Qwen3-TTS and VoxCPM are Apache-2.0 repositories emphasizing voice design and cloning. The sources do not compare any of them with Kokoro.

Quick comparison

How the options differ

Language coverage and specialization

Chatterbox Multilingual V3 is described as a 0.5B general-purpose multilingual model supporting 23 languages. The listed languages include Arabic, Hindi, Japanese, Swahili and Chinese, among others.

Chatterbox also ships a Single Language Pack. It contains dedicated finetunes for Chinese, LatAm Spanish, Brazilian Portuguese, Spain Spanish, Portugal Portuguese and Hindi. The model card positions these for dialect-sensitive applications.

VoxCPM's repository description calls it a tokenizer-free TTS for multilingual speech generation. Its topics include multilingual, voice-cloning and voice-design. The excerpts we read do not list supported languages.

The Qwen3-TTS excerpt does not mention language coverage at all.

Voice control and cloning

Chatterbox documents explicit tuning knobs. The model card says it supports emotion exaggeration control and describes it as the first open source TTS model to do so. It also recommends default settings of exaggeration=0.5 and cfg=0.5, and suggests lowering cfg to around 0.3 for fast-speaking reference voices.

The card warns that the reference clip should match the specified language tag. Otherwise, language transfer outputs may inherit the accent of the reference clip's language.

Qwen3-TTS is described as supporting free-form voice design and vivid voice cloning. VoxCPM is described as supporting creative voice design and true-to-life cloning. Neither excerpt documents parameters comparable to Chatterbox's exaggeration and cfg settings.

Streaming, deployment and safeguards

The Qwen3-TTS repository describes stable, expressive, and streaming speech generation. It is the only one of the three whose excerpt mentions streaming. It is presented as a series of models from the Qwen team at Alibaba Cloud.

Chatterbox publishes a pip install path and Python examples that load the model on a CUDA device. Its card also states that every generated file carries a PerTh neural watermark.

The Chatterbox card mentions a separate hosted service with "sub 200ms" latency. That figure comes from the vendor's own description, and the excerpts give no pricing for it.

License and project age

Chatterbox is stated as MIT licensed. Qwen3-TTS and VoxCPM both report Apache-2.0 through the GitHub API.

The Qwen3-TTS repository was created on 2026-01-21. The VoxCPM repository was created on 2025-09-16. A repository license does not tell you what any hosted inference costs.

Ranking criteria

We ordered the list by how much of the query-relevant detail is documented in the cited sources. We looked at stated license, language coverage, voice-control documentation and usage instructions.

Chatterbox ranks first because its model card documents all four. Qwen3-TTS ranks second because its excerpt states license and streaming. VoxCPM ranks third because its excerpts state license and capabilities but give no language list or usage detail.

This is a ranking of documentation depth, not of audio quality. We do not name a best option.

How we compared

We read the primary sources for each tool on 2026-10-14. For Chatterbox, that was the Hugging Face model page. For Qwen3-TTS, the GitHub repository metadata and description. For VoxCPM, the GitHub repository metadata and the project homepage at voxcpm.com. The homepage excerpt we captured contained only a language-redirect notice, so it adds no product facts.

We did not install or run any of these models. We did not listen to outputs, measure latency or quality, or compare any of them against Kokoro. All statements are attributed to the sources above.

Limitations

• The sources never mention Kokoro, so no head-to-head claims are possible.
• Hardware requirements, model sizes (other than Chatterbox's 0.5B), and inference speed are not stated for Qwen3-TTS or VoxCPM in the cited sources.
• Language lists for Qwen3-TTS and VoxCPM are not stated in the cited sources.
• Hosted pricing is not stated for any tool.
• Chatterbox's benchmark and preference claims are vendor statements, and we did not verify them.

For more, see the voice models category, the hubs for Chatterbox, Qwen3-TTS and VoxCPM, and the related guide Best free AI voice models in 2026.