Best Open Source Text to Speech Models in 2026
Compare the best open source text to speech models in 2026: Kokoro, Chatterbox, Qwen3-TTS, Orpheus, Dia and more, with licenses, languages and hardware needs.
The best open source text to speech model depends on the job. Kokoro-82M is small and Apache-2.0 licensed, Chatterbox clones voices under the MIT license, and Qwen3-TTS covers 10 languages and includes a voice design variant. Before you build anything on one, read the license file, because several popular open-weight models forbid commercial use.
This guide compares the open models that matter most in October 2026, with licenses, languages and documented hardware. Figures come from model cards, READMEs and a leaderboard read on 2026-10-17.
What "open source" means for text to speech
Most "open source" speech models are really open weights. The weights, meaning the trained parameters, can be downloaded, but the license attached decides what you may do with them. Artificial Analysis defines "Open Weights" as "models that can be self-hosted using downloadable model weights or a provider-supplied inference runtime". It tracks "Open Weights Commercial" separately, for models whose published license permits commercial use, and it excludes models that need a separately negotiated commercial agreement.
That gap is the trap. A model can be free to download and still be off limits for a paid product. Code and weights can also carry different licenses. The F5-TTS repository lists MIT for the code, but its pretrained models are licensed CC-BY-NC.
Best open source text to speech models at a glance
Sizes and licenses come from each model's page. The "good first use" column is a summary of what each page emphasizes, not a benchmark result.
| Model | Size | License | Languages | Good first use | |---|---|---|---|---| | Kokoro-82M v1.0 | 82M | Apache-2.0 | 8 on the model card | Small, low-cost self-hosting | | Chatterbox | 500M | MIT | English | Voice cloning from a reference clip | | Chatterbox-Multilingual V3 | 500M | MIT | 23 listed on GitHub | Cloning across many languages | | Chatterbox-Nano | 110M | MIT (repo license) | English | CPU-friendly experiments | | Qwen3-TTS (0.6B, 1.7B) | 0.6B and 1.7B | Apache-2.0 | 10 | Multilingual speech and voice design | | Orpheus TTS | 3B | Apache-2.0 | English main | Expressive English speech | | Dia | 1.6B | Apache-2.0 | English only | Two-speaker dialogue | | Sesame CSM-1B | 1B backbone | Apache-2.0 | English | Context-aware speech |
Kokoro: the smallest model to run yourself
Kokoro-82M is, in its model card's words, "an open-weight TTS model with 82 million parameters". The version 1.0 release was published on 2025-01-27, with 54 voices across 8 languages. It is the easiest entry point on this list. The Python package installs with pip install kokoro>=0.9.4 soundfile and needs espeak-ng for some fallbacks.
Community builds widen the options. The quantized ONNX files range from 86 MB to 326 MB depending on precision, and a community Docker project, Kokoro-FastAPI, exposes an OpenAI-compatible endpoint. The model card states no minimum CPU or GPU, so test on your own machine. The full specs and voice grades are in Kokoro TTS explained.
Chatterbox: voice cloning from a short clip
Chatterbox is MIT licensed and comes from Resemble AI. Its model zoo has several open variants. Chatterbox-Turbo is 350M parameters and English only, with paralinguistic tags such as [laugh], and it needs a reference clip to clone a voice. Chatterbox-Nano is 110M parameters, English only, and runs on CPU. Chatterbox-Multilingual V3 is 500M parameters and lists 23 languages.
Resemble has stopped taking new customers for its hosted voice service. It says the models it built "are open and free to use". The implications for licensing and hosted options are covered in Chatterbox TTS explained.
Qwen3-TTS: multilingual speech with voice design
Qwen3-TTS was released on 2026-01-22 as 0.6B and 1.7B models under Apache-2.0. It supports 10 languages: Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian. The base models offer a "3-second rapid voice clone". Only the 1.7B-VoiceDesign model creates a voice from a natural-language description.
The project page also claims synthesis latency "as low as 97ms". That figure is self-reported, with no hardware or method given, so do not quote it as a general speed.
Orpheus and Dia: expressive speech and dialogue
Orpheus TTS uses a Llama-3b backbone and is Apache-2.0 licensed. Its main model is English. The README lists eight emotion tags (laugh, chuckle, sigh, cough, sniffle, groan, yawn and gasp) and eight built-in English voices. Zero-shot cloning is listed as supported, but the authors say the model is "not explicitly trained for this".
Dia, from Nari Labs, has 1.6B parameters, is Apache-2.0 licensed and handles English only. It uses [S1] and [S2] tags for two-speaker dialogue. The README reports about 4.4 GB of VRAM in bfloat16 or float16 on an RTX 4090.
Sesame CSM: conversational context
Sesame CSM-1B is an Apache-2.0 release of a speech model that uses earlier conversation audio as context. It needs a CUDA GPU, and it is built mainly for English. It is a base model, not the voice in Sesame's demo, and it cannot generate text.
Open weights you cannot use commercially
Several strong models sit behind licenses that block commercial use or require a separate deal. Read the license file before you build on any of them.
- Voxtral TTS (Mistral): CC BY-NC 4.0, which the model card says it inherits. It supports 9 languages and 20 preset voices, and runs on a single GPU with 16 GB or more.
- Breeze TTS 2: BreezeBlue Research and Non-Commercial License. The code is Apache-2.0, but the card says that license "does not grant rights to use the model commercially".
- Fish Audio S2-Pro: Fish Audio Research License. The model page says "Commercial use requires a separate license from Fish Audio".
- VibeVoice: MIT license, but the README says it is "intended for research and development purposes only", and Microsoft removed the VibeVoice-TTS code from the repository on 2025-09-05.
- XTTS-v2: Coqui Public Model License. The model page does not state commercial terms, so read the CPML directly.
How good are they? A dated leaderboard snapshot
Artificial Analysis derives Elo scores from user votes in blind comparisons, where listeners choose which sample sounds more natural. The table below was read on 2026-10-17. Leaderboard numbers move, so always quote them with the date. Ranks are shown as ranges because models with close scores share positions.
| Model | Elo | Rank range | Notes | |---|---|---|---| | Eleven v4 Turbo | 1330 | 1 | Hosted, not open | | Fish Audio Drama 3 Preview | 1244 | 8-10 | Check the model page before use | | Breeze TTS 2 | 1222 | 11-16 | Non-commercial license | | Fish Audio S2 Pro | 1116 | 32-37 | Commercial use needs a separate license | | Voxtral TTS | 1087 | 39-51 | CC BY-NC 4.0 | | Kokoro-82M v1.0 | about 1061 | 52-63 | Apache-2.0 | | Chatterbox | 1025 | 71-78 | MIT |
The reading is straightforward. The top of the arena is hosted, and the strongest open-weights rows carry restrictive licenses. Among the permissively licensed models here, Kokoro and Chatterbox sit lower on naturalness votes. Vendor claims about quality, such as Kokoro's "comparable to larger models", are self-reported.
What you need to run one
| Model | Documented hardware | |---|---| | Kokoro-82M | No minimum stated by the authors. ONNX builds of 86 MB to 326 MB. A community note says kokoro-onnx runs near real time on an M1 Mac | | Chatterbox-Nano | Runs on CPU. The vendor says it is 3x faster than realtime on 8 CPU cores | | Chatterbox (other variants) | CUDA, CPU or Apple MPS. Developed and tested on Python 3.11 | | Sesame CSM-1B | CUDA-compatible GPU. No official memory figure | | Dia | About 4.4 GB of VRAM in bfloat16 or float16 on an RTX 4090 | | Voxtral TTS | Single GPU with 16 GB or more | | Breeze TTS 2 | About 7.7 GiB of GPU memory in eager mode, with 12 GB recommended |
A laptop CPU is a realistic target for Kokoro and Chatterbox-Nano. For the other models on this list, plan on a GPU.
When a hosted tool makes more sense
Running an open model means paying for GPU time, checking licenses and keeping the setup working. If you want speech without that work, a hosted studio is simpler. OndaLabs is a hosted AI voice studio with text to speech, two-speaker dialogue, speech to text, custom voices, image generation and music generation. It offers 30 prebuilt voices usable in every supported language, and 37 languages have dedicated pages.
The free plan is $0 with 2,000 credits per month, about 2 minutes of speech. Commercial use is included from the Starter plan at $9. OndaLabs is hosted and is not open source, so it cannot be self-hosted. For the cost math behind that choice, read self hosted text to speech vs hosted APIs. You can compare plans on the pricing page or try text to speech directly.
Frequently asked questions
What is the best open source text to speech model in 2026?
There is no single winner. Kokoro-82M suits small, low-cost self-hosting, Chatterbox suits voice cloning under the MIT license, and Qwen3-TTS suits multilingual work and voice design. The top open-weights models on the Artificial Analysis snapshot score higher, but they carry non-commercial or restricted licenses.
Which open source TTS models can I use commercially?
Models whose pages state Apache-2.0 or MIT, such as Kokoro-82M, Chatterbox, Qwen3-TTS, Orpheus and Dia, are permissive on paper. Sesame CSM-1B is also Apache-2.0, but its model card says it is intended for research and educational purposes, so read that card before shipping. Always read the license file, because code and weights can differ.
Can open source TTS clone a voice?
Some models can. Chatterbox clones from a reference clip, and the Qwen3-TTS base models offer a "3-second rapid voice clone". Kokoro documents no cloning feature at all. Only clone a voice you have permission to use.
What is the best open source TTS that runs on a CPU or laptop?
Kokoro has a community CPU Docker image and ONNX builds, and Chatterbox-Nano runs on CPU according to its README. The authors of Kokoro do not publish a minimum CPU, so test on your own hardware. Speed figures on these pages come from vendors or community testers and depend heavily on the machine.
Is open source TTS as good as ElevenLabs?
It depends on the voice and the task, and no single test settles it. In the Artificial Analysis snapshot read on 2026-10-17, the top hosted model scored 1330, while the Apache-2.0 Kokoro model scored about 1061. Treat both numbers as a dated snapshot of listener votes, not a permanent ranking.
What is the difference between open source and open weights?
Open weights means the trained parameters can be downloaded and run. Open source usually refers to code and a license that grants permission to use, modify and share it. Many TTS models are open weights with non-commercial terms, which is why the license matters more than the label.