Open-source text to speech has gotten good. An 82-million-parameter model now scores within a few points of Amazon Polly’s generative voices in blind listening tests, and you can download it for free. The catch is in the word “open.” Three of the thirteen models in the table below come with weights you can’t use in anything you sell, a fourth asks you not to, and several popular ones not in the table have the same problem.
So this guide sorts open-source TTS models by license first and quality second. It’s for developers deciding what to put in an app, a device, or a voice agent. If you’re building something for yourself, skip the license column and pick whatever sounds best.
The short answer
| Model | Who makes it | Weights license | Commercial use | Languages | Artificial Analysis score |
|---|---|---|---|---|---|
| Kokoro 82M | hexgrad | Apache 2.0 | Yes | 8 | 1064 |
| Magpie TTS Multilingual | NVIDIA | NVIDIA Open Model License | Yes | 12 | 1064 |
| Maya1 | Maya Research | Apache 2.0 | Yes | English | 1046 |
| Chatterbox | Resemble AI | MIT | Yes, output watermarked | 23+ (Multilingual V3) | 1026 (English model) |
| Zonos v0.1 | Zyphra | Apache 2.0 | Yes | 5 | 1000 |
| Orpheus 3B | Canopy Labs | Apache 2.0 | Yes | English | Not rated |
| Dia 1.6B | Nari Labs | Apache 2.0 | Yes | English | Not rated |
| CosyVoice 2 | Alibaba | Apache 2.0 | Yes | Chinese-first, multilingual | Not rated |
| Piper | Open Home Foundation | GPL-3.0 engine, per-voice licenses | Check each voice | 40+ locales | Not rated |
| VibeVoice 1.5B | Microsoft | MIT, research use only | Not recommended | English, Chinese | 957 |
| XTTS v2 | Coqui (closed) | Coqui Public Model License | No | 17 | 916 |
| Fish Audio S2 Pro | Fish Audio | Fish Audio Research License | No | 80+ | 1116 |
| Breeze TTS 2 | BreezeBlue | Research and non-commercial | No | English, Chinese | 1220 |
Scores are Elo ratings from the Artificial Analysis Provider Voice leaderboard, October 2026. Listeners hear two clips of the same line without knowing which model made either and pick the one they prefer. Licenses and language counts come from each model’s own card or repository, checked the same month. “Not rated” means the model isn’t on that board, not that it’s bad.
If you want one pick: Kokoro for small and fast English, Chatterbox for voice cloning and many languages, Piper for offline devices.
Check the weights license before anything else
An open-source TTS project usually has two licenses: one for the code and one for the trained weights. The code is almost always permissive. The weights are where the restrictions live, and without the weights the code makes no sound.
XTTS v2 is the classic example. The Coqui TTS library is MPL-2.0 and still maintained in a fork by the Idiap Research Institute. The XTTS v2 weights ship under the Coqui Public Model License, which opens with: “This license allows only non-commercial use of a machine learning model and its outputs.” The outputs part matters: the audio you generate inherits the restriction. Coqui used to sell commercial licenses. The company has since shut down, so that door is closed.
The same pattern shows up across the newer models. F5-TTS has MIT code and CC BY-NC 4.0 weights. Mistral says Voxtral TTS inherits CC BY-NC 4.0 from the reference voices it ships with. Fish Audio’s S2 Pro uses the Fish Audio Research License, and its model card asks you to agree to non-commercial use only. Breeze TTS 2, the highest-rated open-weights model on Artificial Analysis, licenses its weights and even self-hosted outputs for research and non-commercial use only.
VibeVoice is a special case. Microsoft released it under MIT, which is as permissive as licenses get, but the model card limits it to research and says Microsoft doesn’t recommend it for commercial or real-world use without further testing. Microsoft also disabled the weights for the larger VibeVoice-Large model. An MIT license with a “please don’t” attached is a legal question for your lawyer, not a blog post.
The models you can ship
Kokoro 82M is the default answer for English. It has 82 million parameters, an Apache 2.0 license, and a score of 1064, level with NVIDIA’s Magpie and two points behind Amazon Polly’s generative voices. Version 1.0 ships 54 voices across 8 languages. It’s small enough that hosted providers were reselling it for under $1 per million characters by April 2025, according to its model card. It doesn’t clone voices.
Chatterbox from Resemble AI is the pick if you need voice cloning under a permissive license. The MIT-licensed family includes a 500M English model, Chatterbox Multilingual V3 at 500M and 23+ languages, and Chatterbox-Turbo, a 350M English model built for lower latency, with tags such as [laugh] and [cough]. Every file Chatterbox generates carries Resemble’s Perth watermark, which is invisible to listeners but detectable. That’s a feature if you care about provenance and a surprise if you didn’t read the card.
NVIDIA Magpie TTS Multilingual covers 12 languages, including Hindi, Japanese, Korean and Arabic, with 364M parameters. It runs on NVIDIA’s NeMo framework. The NVIDIA Open Model License says plainly that the models are commercially usable. It ships five English speaker voices rather than a big catalog.
Orpheus 3B from Canopy Labs is built on a Llama 3B backbone and reports about 200ms streaming latency, or about 100ms with input streaming. The weights are listed as Apache 2.0; because the backbone is Llama, read Meta’s Llama license alongside it. Its two main models are English, and the multilingual models are a research preview.
Dia 1.6B writes dialogue, not narration. You tag speakers with [S1] and [S2], and it can add nonverbal sounds such as laughter. It’s Apache 2.0 and English only. Without a reference clip it picks a different voice on every run, which is charming in a demo and less so in a product.
CosyVoice 2 from Alibaba’s FunAudioLLM team is Apache 2.0 and supports streaming both text in and audio out, with latency as low as 150ms by its own account. Its language coverage centers on Chinese and its dialects.
Zonos v0.1 from Zyphra (Apache 2.0) speaks English, Japanese, Chinese, French and German and needs 6GB or more of GPU memory. Maya1 (Apache 2.0, 3B parameters, English) is built for expressive voice design, with tags for laughing, crying and whispering.
Piper is the odd one out: a fast, local neural TTS engine from the Open Home Foundation, the group behind Home Assistant. The maintained engine, piper1-gpl, is GPL-3.0. If you embed it, GPL obligations follow. Each voice has its own model card and training-data license, so check the voice, not just the engine.
How good is open-source TTS, honestly
Good enough that you’d use it on purpose. The best permissively licensed models score in the 1000 to 1065 range on Artificial Analysis. That’s the neighborhood of Amazon Polly’s generative voices (1066) and Google’s Chirp 3 HD (1059).
The ceiling is higher. Breeze TTS 2 reaches 1220 and Fish Audio S2 Pro 1116, but neither is licensed for commercial use. Cartesia’s Sonic 3.6 scores 1281, about 200 points above Kokoro. On an Elo scale, a 200-point gap means listeners prefer the higher-rated model in about three out of four head-to-head comparisons.
A leaderboard also measures one thing: which clip people prefer, heard in isolation. It doesn’t test how a model reads “your confirmation code is Q7-4XB,” whether it keeps the same voice across a ten-minute call, or how fast the first audio arrives when 200 calls start at once. We wrote about why one score can’t tell you whether a TTS model is good. Run your own scripts, in your own languages, through the path your users will hear.
Where self-hosting gets expensive
Downloading weights is free. Running them for real users isn’t.
For a voice agent on the phone, the model is the easy part. You also need GPUs sized for your peak concurrency, a streaming server that starts playback before the sentence is finished, text normalization so “$1,250.00” and “Dr.” are read correctly, and failover when a node dies mid-call. Several of these models ship as research code that generates a whole clip before returning any of it. Turning that into sub-second streaming under load is an engineering project, and it’s the project the model card doesn’t mention.
That trade is worth it in some cases:
- You need to run offline, on a device, or inside an air gap, and your quality bar fits a small model.
- Your volume is high, steady and predictable, so owned GPUs beat per-character pricing.
- You’re building something for yourself and the license doesn’t matter.
If your reason is data control rather than cost, you may not need open weights at all. Hosted providers, Cartesia included, offer on-premise and VPC deployments where the audio never leaves your network.
When to use Cartesia instead
Cartesia’s Sonic is not open weights. It’s a hosted, streaming text to speech model built for voice agents: sub-90ms time to first audio, 44 languages from one model, and instant voice cloning. The free plan includes 20,000 credits a month, and one credit is one character, so you can test it on your own scripts before paying anything. For regulated deployments, the same models run on-premise or air-gapped under an enterprise contract.
Our open-source work is Edge, an Apache 2.0 library for running state space models on-device. If you need a free, offline English voice for a hobby project, Kokoro or Piper is the better answer, and we’d tell you that in person too.