Open Source ElevenLabs Alternatives: The License Verdict Table for Voice Models
By the SoftSea editorial team
Photo: Tima Miroshnichenko / Pexels
Open source is a verdict, not a vibe
Voice AI has a labeling problem: a model can be downloadable, inspectable, and famous, and still forbid the commercial use its roundup implies. The incumbents for this query (localaimaster, oakgen's October 25, 2026 license table, taku, fdaytalk) increasingly print licenses, which is progress, but they print one license per model, and modern TTS has two: the code that runs the model and the weights that are the model. This guide prints both, plus the only column that decides anything: may you sell what it says. The GitHub companion guide grades the repositories themselves; the ranked alternatives guide covers the hosted field.
Model
Code license
Weights license
Commercial verdict
Cloning
Chatterbox
MIT
MIT release, PerTh watermark
Yes
Yes, zero-shot
Kokoro-82M
Apache-2.0
Apache-2.0
Yes
No
OpenVoice v2
MIT
MIT
Yes
Yes, tone cloning
GPT-SoVITS
MIT
MIT release
Yes
Yes, few-shot
Piper
GPL-3.0
Voices vary by voice
Yes, under GPL terms
No
F5-TTS
MIT
CC-BY-NC
No, not with the released weights
Yes
XTTS v2
MPL-2.0 toolkit
Coqui Public Model License
No
Yes
Fish Speech
Community license (source-available)
Community license
Restricted; not open source
Yes
The clean licenses: Chatterbox, Kokoro, OpenVoice, GPT-SoVITS
Chatterbox, from Resemble AI, is the strongest open cloning release available: MIT in both layers, zero-shot cloning from a few seconds of reference audio, 23 languages, and a built-in Perth watermark that embeds provenance into generated audio. Treat the watermark as part of the deal: it is the project's answer to misuse, and a commercial user should disclose rather than strip it. Kokoro-82M is the narration counterpart: Apache-2.0 in both layers, 54 voices across 8 languages, no cloning, and a quality-to-size ratio that embarrasses models fifty times larger; it runs in real time on an ordinary CPU, which makes it the default engine for apps, devices, and unlimited local narration. OpenVoice v2 (MIT) specializes in tone-color cloning with a base speaker model, a good fit when style transfer rather than identity cloning is the job. GPT-SoVITS (MIT) is the tinker's favorite for few-shot cloning with fine-tuning, powerful and correspondingly demanding of its operator. These four are what "open source ElevenLabs alternative" means when the words are used precisely: download, run, sell the output, owe nothing but the license notice.
Interactive tool
Comparison Builder
Pick two or three products and see them side by side on the numbers that decide the purchase.
Measure
Chatterbox
Kokoro-82M
Piper
Starting paid price
Free
Free
Free
Free option
Free plan: fully free and open source (MIT)
Free plan: fully free and open source (Apache-2.0)
Free plan: fully free and open source (GPL-3.0)
Plans
Chatterbox: free
Kokoro-82M: free
Piper: free
Platforms
Windows, macOS, Linux
Windows, macOS, Linux
Windows, macOS, Linux
Best for
developers cloning voices, local-first production, research
home automation, offline devices, privacy-first local speech
Strengths
Real zero-shot cloning, self-hosted; MIT license allows commercial use; Runs locally; audio never leaves your machine
Runs on a laptop CPU in real time; Apache-2.0: commercial use is clean; Quality punches far above its size
Offline and very fast on small hardware; The default voice engine in Home Assistant; Huge community voice catalog
Watch out for
You supply the GPU and the setup time; Watermarking is part of the design; No hosted support or SLA
No voice cloning at all; Voice count and languages are fixed; Assembly into a product is on you
Voices are functional, not studio-grade; No cloning; Original repository was archived in October 2025; development moved to piper1-gpl
Lowest entry price: Chatterbox at Free per month flat. Real free plan: Chatterbox, Kokoro-82M, Piper. Entry price is the floor, not the decision: check the plan line above for the tier that actually includes the feature you need.
All figures from vendor pricing pages, checked November 18, 2026. Some vendors price by contacts, tasks, or operations instead of seats; those notes are in the plan data and the linked pricing guides.
The conditional: Piper and GPL
Piper is GPL-3.0, which permits commercial use and even selling products, with copyleft obligations when you distribute the software itself. For a service that merely uses Piper internally to render audio, the obligations are minimal; for a device or app that ships Piper inside, the GPL's source obligations apply to the distributed code. Its development moved to piper1-gpl under the Open Home Foundation after the original repository was archived in October 2025. Piper's voices are functional rather than expressive; its kingdom is offline, embedded, and home automation speech, where it has no serious open rival.
The corrections: XTTS, F5-TTS, Fish Speech
XTTS v2 is an excellent cloner under the Coqui Public Model License, which forbids commercial use. It is not open source for business purposes, full stop, and Coqui's shutdown means no commercial license is coming. F5-TTS is subtler: MIT code, CC-BY-NC pretrained weights. The repository is open source; the model everyone actually downloads is non-commercial, and using it commercially means training or obtaining weights under different terms, a project, not a download. Fish Speech is source-available under its own community license: you may read and run it under conditions, which is a legitimate choice by its authors and is not open source under any standard definition. None of these are bad projects. All three are mislabeled by lists that equate "on GitHub" with "open", and the mislabeling is the risk: a product built on NC weights discovers its license at the worst possible moment, which is after launch.
Interactive tool
Open source license audit
Check off each step as you finish it. Your progress stays on this page.
0 of 6 complete
0% done
Running them: the honest hardware paragraph
Open models trade subscription bills for hardware and assembly. Kokoro and Piper run on CPUs and cost nothing to operate at any volume. Chatterbox, OpenVoice, GPT-SoVITS, and F5-TTS want a GPU for comfortable speed, which means a local card or a rented instance whose hourly rate belongs in your unit economics (the cheaper companion guide shows the API prices you are competing with, down to Polly at $4 per million characters). The verdict: if your volume is high and your content is yours, the open bench is not the cheap option, it is the sovereign one: no metering, no terms change, no vendor between your manuscript and your microphone.
Related guides
Before you commit, it is worth a look at the nearest alternatives and comparisons, gathered below from the same section of the chart.
Which open source model is best for voice cloning?
Chatterbox: MIT in both code and weights, zero-shot cloning, 23 languages, actively maintained, with a built-in watermark to disclose. GPT-SoVITS (MIT) is the alternative for few-shot fine-tuning workflows. Avoid building commercial cloning on XTTS (non-commercial license) or F5-TTS's released weights (CC-BY-NC).
Is XTTS v2 open source?
Its toolkit is MPL-2.0, but the XTTS v2 model is under the Coqui Public Model License, which prohibits commercial use. For business purposes it is not open source, and since Coqui has shut down, no commercial license is available. It remains free for non-commercial use.
What is the best open source model for narration without cloning?
Kokoro-82M: Apache-2.0 for code and weights, 54 voices, real-time on a CPU, and quality well above its size. It is the easiest model in this guide to embed in an app or a local pipeline with zero licensing friction and zero per-character cost.
Is Fish Speech open source?
No. Fish Speech is source-available under a community license with restrictions. You can inspect and run it under its terms, but it does not meet the open source definition, and commercial users must read its license rather than a roundup's label.
Do open source voices match ElevenLabs quality?
The best open cloners (Chatterbox class) are close enough that many listeners cannot tell in short clips; ElevenLabs still leads on expressive range, long-form consistency, and zero setup. The open bench wins on license, sovereignty, and marginal cost, which for high-volume or ownership-sensitive work matters more than the last few percent of polish.
Photo credits: Tima Miroshnichenko (Pexels). Photos are served from this site; follow the credit to the library page for the original.