The open-source alternatives to ElevenLabs are Chatterbox, Kokoro, Piper, Fish Speech and F5-TTS, plus two runtimes, Speaches and sherpa-onnx, that serve such models through an API or on a phone. Chatterbox and Kokoro carry MIT and Apache licences on both code and weights, so they are the ones a company can ship with. Piper is the small, fast engine behind Home Assistant. Fish Speech and F5-TTS produce voice clones from a ten-second clip, and both restrict commercial use of their weights, which this page spells out. A team switches when it wants the audio generated on its own hardware, no per-character bill, and a licence it has read.
10 of 13 carry an OSI licence. 13 of 13 run on your own machines. Figures read from GitHub on October 26, 2026.
One project per need, with the reason from its own README and the one thing to know before you choose it.
MIT code and MIT weights, cloning from a reference clip of about ten seconds, and a multilingual model covering 23 languages. The one cloner here whose licence needs no negotiation.
Know firstStreaming is not described in the README, and the tested platform is Debian 11.
Apache-2.0 code and weights around the 82-million-parameter Kokoro model, served through an OpenAI-compatible endpoint with streaming, in Docker images for CPU and NVIDIA.
Know firstPreset voices only. The README calls its voice mixing a tuner, not a cloner.
The fast local engine behind Home Assistant and the NVDA screen reader, with a pip install, a web server and a C library.
Know firstNo cloning, voice licences vary by voice, and the foundation behind it is looking for maintainers.
Cloning from a 10 to 30 second clip, inline emotion tags, over 80 languages by the project's count, and streaming through SGLang.
Know firstCode and weights are under a research licence; any commercial use needs a separate agreement with Fish Audio.
MIT code with zero-shot cloning, a Gradio app, and documented training and fine-tuning for a checkpoint on audio you have the rights to.
Know firstThe published weights are CC-BY-NC-4.0 because of their training data, so they cannot ship in a product.
MIT. One OpenAI-compatible server that serves Kokoro and Piper voices and transcribes with Whisper, loading models on demand.
Know firstA short README and a release-candidate tag; the models it loads carry their own licences.
AGPL-3.0 workflow engine over 16 speech engines and 10 transcription engines, with a local API and an MCP endpoint for agents, all on your own hardware.
Know firstThe default OmniVoice engine's weights are CC-BY-NC; the README says to review each engine's licence before commercial use.
Tencent Hunyuan's 1.5-billion-parameter model under MIT for code and weights, covering zero-shot TTS plus content, emotion, pitch and timbre editing, enhancement and separation.
Know firstA research release for an NVIDIA GPU, with no packaged app, no streaming API and no tagged release in the README.
Open-weight bilingual model with cloning, voice design and inline vocal events; the README quotes under 40 milliseconds to first audio on an H100.
Know firstThe weights and every self-hosted output are under a research and non-commercial licence; commercial use goes through the paid hosted API.
Apache-2.0 runtime that runs Piper, Matcha and voice-cloning models on CPUs, NPUs, Android, iOS and WebAssembly with no server.
Know firstA lower-level tool; you choose the models and read their licences.
The facts that decide most choices. Each project's page carries the full panel.
| Project | Best for | Openness | Licence | Runs as | Own machines | Stars |
|---|---|---|---|---|---|---|
| Chatterbox | Voice cloning you can ship | OSI | MIT | pip, Python library | Yes | 27,280 |
| Kokoro FastAPI | A speech API on your own server | OSI | Apache-2.0 | Docker, Server (OpenAI-compatible API) | Yes | 5,596 |
| Piper | Small devices and offline assistants | OSI | GPL-3.0 | pip, CLI, Server, C/C++ library | Yes | 6,034 |
| Fish Speech | Expressive cloning, non-commercial | Source available | See LICENSE | Docker, CLI, Web UI, Server | Yes | 33,408 |
| F5-TTS | Research, or training your own weights | Open weights | MIT | pip, Docker, CLI, Web UI (Gradio) | Yes | 15,458 |
| Speaches | Speech in and speech out from one container | OSI | MIT | Docker, Server (OpenAI-compatible API) | Yes | 3,736 |
| VoiceStudio | Cloning, dubbing and design in one desktop app | OSI | AGPL-3.0 | Desktop app (Electron), Docker, Local API, MCP server | Yes | 88,676 |
| AuK | Editing existing speech, not only generating it | OSI | See LICENSE | pip, Python API, Gradio web app, ComfyUI nodes | Yes | 1,548 |
| Breeze TTS 2 | Real-time English and Chinese, non-commercial | Open weights | Apache-2.0 | pip, Docker, CLI, Server (streaming API) | Yes | 945 |
| sherpa-onnx | Speech inside a phone or embedded app | OSI | Apache-2.0 | Libraries in 12 languages, prebuilt Android and Flutter apps, WebAssembly, websocket server | Yes | 15,528 |
| LocalAI | OSI | MIT | Docker, macOS app, Kubernetes | Yes | 49,980 | |
| Vanta | OSI | See LICENSE | Node.js project on top of Remotion | Yes | 155 | |
| Simpaudio | OSI | MIT | Desktop app (installer, portable, or Python) | Yes | 13 |
Licence and star figures read from GitHub on October 26, 2026. Stars are shown as a dated fact and were not used to rank this page.
A family of text-to-speech models from Resemble AI under MIT, with zero-shot voice cloning from a reference clip and a multilingual model covering 23 languages.
MIT code and MIT weights, cloning from a reference clip of about ten seconds, and a multilingual model covering 23 languages. The one cloner here whose licence needs no negotiation.
Know firstStreaming is not described in the README, and the tested platform is Debian 11.
A Docker image that serves the 82-million-parameter Kokoro model through an OpenAI-compatible speech endpoint, with streaming, voice mixing and nine languages, all under Apache-2.0.
Apache-2.0 code and weights around the 82-million-parameter Kokoro model, served through an OpenAI-compatible endpoint with streaming, in Docker images for CPU and NVIDIA.
Know firstPreset voices only. The README calls its voice mixing a tuner, not a cloner.
A fast, local neural text-to-speech engine maintained by the Open Home Foundation, used by Home Assistant and the NVDA screen reader, under GPL-3.0.
The fast local engine behind Home Assistant and the NVDA screen reader, with a pip install, a web server and a C library.
Know firstNo cloning, voice licences vary by voice, and the foundation behind it is looking for maintainers.
A multilingual text-to-speech system with voice cloning from a 10 to 30 second clip and inline emotion tags, released with its weights under a research licence that requires a separate agreement for commercial use.
Cloning from a 10 to 30 second clip, inline emotion tags, over 80 languages by the project's count, and streaming through SGLang.
Know firstCode and weights are under a research licence; any commercial use needs a separate agreement with Fish Audio.
A research text-to-speech model with zero-shot voice cloning whose code is MIT but whose pretrained weights are non-commercial, because of the dataset they were trained on.
MIT code with zero-shot cloning, a Gradio app, and documented training and fine-tuning for a checkpoint on audio you have the rights to.
Know firstThe published weights are CC-BY-NC-4.0 because of their training data, so they cannot ship in a product.
An OpenAI-compatible server under MIT for streaming transcription, translation and text to speech, loading Whisper, Kokoro and Piper models on demand, described by its maintainers as Ollama for speech models.
MIT. One OpenAI-compatible server that serves Kokoro and Piper voices and transcribes with Whisper, loading models on demand.
Know firstA short README and a release-candidate tag; the models it loads carry their own licences.
An AGPL-3.0 desktop workflow engine for voice cloning, voice design, video dubbing, dictation and audiobooks, running 16 speech engines and 10 transcription engines locally, with a local API and an MCP server for agents.
AGPL-3.0 workflow engine over 16 speech engines and 10 transcription engines, with a local API and an MCP endpoint for agents, all on your own hardware.
Know firstThe default OmniVoice engine's weights are CC-BY-NC; the README says to review each engine's licence before commercial use.
Tencent Hunyuan's 1.5-billion-parameter speech foundation model under MIT for code and weights, covering zero-shot text to speech, speech editing, enhancement and separation through one natural-language instruction interface.
Tencent Hunyuan's 1.5-billion-parameter model under MIT for code and weights, covering zero-shot TTS plus content, emotion, pitch and timbre editing, enhancement and separation.
Know firstA research release for an NVIDIA GPU, with no packaged app, no streaming API and no tagged release in the README.
An open-weight bilingual English and Chinese text-to-speech model built for real-time interaction, with voice cloning, voice design and inline vocal events, whose weights and self-hosted outputs are non-commercial.
Open-weight bilingual model with cloning, voice design and inline vocal events; the README quotes under 40 milliseconds to first audio on an H100.
Know firstThe weights and every self-hosted output are under a research and non-commercial licence; commercial use goes through the paid hosted API.
A runtime under Apache-2.0 from the next-generation Kaldi team that runs speech to text, text to speech, speaker diarization and voice activity detection locally on CPUs, NPUs and phones, with no internet connection.
Apache-2.0 runtime that runs Piper, Matcha and voice-cloning models on CPUs, NPUs, Android, iOS and WebAssembly with no server.
Know firstA lower-level tool; you choose the models and read their licences.
An MIT engine that runs text, vision, voice, image and video models on any hardware, no GPU required, behind drop-in OpenAI, Anthropic and ElevenLabs-compatible APIs, over 60+ backends including llama.cpp, vLLM and whisper.cpp.
A programmatic AI video engine built on Remotion under MIT, that wires 40+ open-source voice, avatar, caption, video, music and image projects into one local render pipeline.
An MIT-licensed Windows text-to-speech studio that runs Piper and Kokoro locally for voice synthesis, Whisper for speech-to-text, with voice blending, a visual SSML editor and SRT export.
The questions that settle it, in the order they usually come up.
Chatterbox (MIT) or Kokoro (Apache-2.0). Fish Speech and F5-TTS both restrict commercial use of their weights, so they stay in research and evaluation.
Chatterbox for commercial use. Fish Speech or F5-TTS for non-commercial use, with the closest clones by their own demos. Kokoro and Piper do not clone; they ship preset voices.
Without one, Kokoro FastAPI ships CPU images and quotes under a second to first audio on an M3 Pro, Piper is built for small machines, and Chatterbox Nano is described as running faster than real time on 8 CPU cores.
Kokoro FastAPI or Speaches. Both present that API shape, so the change is a base URL.
ElevenLabs, in the vendor's own words, is an AI voice generator and voice agents platform with access to 5,000+ voices in 70+ languages through APIs and SDKs, by the vendor's own count. elevenlabs.io. The cases below are where that still wins.
Stars, licence, last commit and last release are read from the GitHub API by a script every week and carry the date they were read. Nobody types them.
What each project does is written from its own README and LICENSE, read in full on the date shown, and phrased as the project's claim. The sources are listed at the end.
That a project is faster, better or cheaper than ElevenLabs. A price for anyone. A user count. A roadmap. If a fact is not on this page, it was not in the source.
It depends on what you need to ship. For a commercial product, Chatterbox (MIT code and weights, voice cloning from a reference clip) and Kokoro (Apache-2.0 code and weights, preset voices, 82 million parameters) are the two whose licences allow it without a separate agreement. For a small device or a home automation setup, Piper is the one Home Assistant uses. For the closest voice clone, Fish Speech and F5-TTS, both of which restrict commercial use.
Yes. Chatterbox, Kokoro, Piper, Fish Speech and F5-TTS all run on your own machine with no account and no network call once the model is downloaded. Kokoro FastAPI and Piper both run on a CPU. Chatterbox Nano is described by its maintainers as running faster than real time on 8 CPU cores.
Chatterbox, Fish Speech and F5-TTS all clone a voice from a short reference recording, typically ten to thirty seconds, with no training step. Kokoro does not clone voices: it ships preset voices and a mixing feature its own README calls a tuner rather than a cloner. Piper uses trained voice models rather than cloning.
Chatterbox (MIT), Kokoro (Apache-2.0) and Piper (GPL-3.0 on the code; check the licence of each voice model separately). Fish Speech releases code and weights under the Fish Audio Research License, which requires a separate written licence for any commercial purpose. F5-TTS code is MIT but its pretrained weights are CC-BY-NC-4.0, which excludes commercial use.
This page does not rank voice quality, because nobody here has measured it on a shared test. Each project's demo page is linked from its own page so you can listen with your own text. What can be stated: Chatterbox and Fish Speech publish inline emotion tags, Kokoro is a small model that runs on a CPU, and Piper is built for speed on small devices.
Kokoro FastAPI exposes an OpenAI-compatible speech endpoint on port 8880 that works with the OpenAI SDKs. Speaches, listed under speech-to-text engines, serves Kokoro and Piper voices through the same OpenAI-compatible shape. Chatterbox and F5-TTS are Python libraries with a Gradio web app, so a small wrapper is needed for an HTTP API.
Open source means the code and the weights both carry a licence on the OSI approved list, such as MIT or Apache-2.0. Open weights means you can download and run the model, but the training data or code is closed, or the weights carry a restriction the code does not. F5-TTS is the clearest example on this page: MIT code, non-commercial weights.
Reveneau sets up the replacement on your own servers, connects it to your product, and builds the software around it.