Open-source alternatives to ElevenLabs: 4 text-to-speech tools compared
Replacing a hosted voice API means running the model on your own hardware and living with the licence the weights come under, and on this page that matters more than quality. Chatterbox-TTS-Server is the one most readers should look at first, because it is the only candidate that is both an HTTP service with an OpenAI-compatible API and permissively licensed.
What ElevenLabs does, and why people look elsewhere
ElevenLabs is a hosted text-to-speech and voice-cloning API with a web studio for narration, dubbing and conversational agents.
- pricing by characters generated
- voice clones and audio are processed on the vendor's servers
- no offline use
The candidates at a glance
| Project | Licence | Maintenance | Language | GitHub stars | Install | Best for |
|---|---|---|---|---|---|---|
| coqui-ai/TTS | MPL-2.0Weak copyleft: changes to its own files stay open | No commits for over a yearLast push August 16, 2024 | Python | 46,103 | See analysis | A Python team that wants a library rather than a service: synthesis, voice cloning from a short reference clip, fine-tuning recipes for your own dataset, and a Trainer API, all without leaving the Python ecosystem. The README lists pretrained models in more than a thousand languages. |
| OHF-Voice/piper1-gpl | GPL-3.0Copyleft: distributing it means sharing source | Commits in the last six monthsLast push October 11, 2026 | C++ | 5,825 | Command documented | Offline narration, accessibility output and embedded devices where a small, fast engine matters more than expressiveness. It installs as piper-tts and ships a Python API, a C and C++ library, an HTTP server and a CLI, with a published set of voices. |
| fishaudio/fish-speech | Custom licenceCustom licence: read the LICENSE file | Commits in the last six monthsLast push September 16, 2026 | Python | 32,942 | See analysis | Multilingual pipelines where prosody and emotion matter and you want explicit control over them: the README describes sub-word level control through natural language tags in the text, multi-speaker and multi-turn generation, and a 4B Dual-Autoregressive model trained on more than ten million hours of audio. |
| devnen/Chatterbox-TTS-Server | MITPermissive: commercial use allowed | Commits in the last six monthsLast push May 26, 2026 | Python | 1,461 | Command documented | A team that wants an HTTP service with a web UI and OpenAI-compatible endpoints without writing the service layer, and who can pin Python 3.10 or use the container images. The README describes the Original, Multilingual and Turbo models, 23 languages, voice cloning, chunking for large text and audiobook generation. |
Each alternative in detail
coqui-ai/TTS🐸💬 - a deep learning toolkit for Text-to-Speech, battle-tested in research and production
- Best for
- A Python team that wants a library rather than a service: synthesis, voice cloning from a short reference clip, fine-tuning recipes for your own dataset, and a Trainer API, all without leaving the Python ecosystem. The README lists pretrained models in more than a thousand languages.
- What you give up
- The cadence is the issue. The recorded last push is 2024-08-16, the newest listed release is v0.22.0 from 2023-12-12, and our analysis of Coqui TTS tells you to treat it as a stable codebase rather than a moving target and to skip it if you need regular releases or a hosted uptime commitment. It is a library, so an HTTP API, a web studio and multi-speaker management are yours to build.
OHF-Voice/piper1-gplFast and local neural text-to-speech engine
- Best for
- Offline narration, accessibility output and embedded devices where a small, fast engine matters more than expressiveness. It installs as piper-tts and ships a Python API, a C and C++ library, an HTTP server and a CLI, with a published set of voices.
- What you give up
- The licence is GPL-3.0, and our analysis of Piper1 GPL says to skip it if you cannot accept copyleft terms. Pronunciation comes from espeak-ng, which the README says is embedded for phonemization, so the voice sounds like a trained voice reading a phonemised sentence rather than a cloned speaker. The README is also currently asking for maintainers, which is worth weighing before you depend on it.
- First install step
- Full setup guide
pip install piper-tts
fishaudio/fish-speechSOTA Open Source TTS
- Best for
- Multilingual pipelines where prosody and emotion matter and you want explicit control over them: the README describes sub-word level control through natural language tags in the text, multi-speaker and multi-turn generation, and a 4B Dual-Autoregressive model trained on more than ten million hours of audio.
- What you give up
- The weights and code are under the FISH AUDIO RESEARCH LICENSE, which our analysis of Fish Speech says is not a standard open-source licence and needs review before any commercial deployment, so this is the one candidate where the licence is the first question rather than the last. Our analysis records no install command for it, and a 4B model is a real infrastructure commitment compared with the other three.
devnen/Chatterbox-TTS-ServerSelf-host the powerful Chatterbox TTS model. This server offers a user-friendly Web UI, flexible API endpoints (incl. OpenAI compatible), predefined voices, voice cloning, and large audiobook-scale text processing. Runs accelerated on NVIDIA (CUDA), AMD (ROCm), and CPU.
- Best for
- A team that wants an HTTP service with a web UI and OpenAI-compatible endpoints without writing the service layer, and who can pin Python 3.10 or use the container images. The README describes the Original, Multilingual and Turbo models, 23 languages, voice cloning, chunking for large text and audiobook generation.
- What you give up
- It is a single-maintainer project, and our analysis of Chatterbox-TTS-Server says to skip it if you need a published support window; the last push was 2026-05-26. The MIT licence covers the server, not the model, so that analysis also tells you to check the voice cloning consent requirements in the upstream model licence. The install also pins extra packages, including an onnx build and a tokenizer, which is a dependency surface to track.
- First install step
- Full setup guide
pip install -r requirements.txt pip install --no-deps git+https://github.com/devnen/chatterbox-v2.git@master s3tokenizer==0.3.0 onnx==1.16.0
What a hosted voice API gives you that a local model does not
ElevenLabs is a hosted text-to-speech and voice-cloning API with a web studio for narration, dubbing and conversational agents. Three documented things send people looking for something else. Pricing is counted by characters generated. Voice clones and the audio itself are processed on the vendor's servers. And there is no offline use.
A local model removes all three. What you give up is not only the cost model but the parts of the product that are not a model: the studio, the dubbing timeline, the conversational agent plumbing, the latency engineering, and the consistency work that keeps one voice sounding like the same person across a long file. Those are engineering problems, and on this page they are handled in three different ways. One project is a library and leaves all of them to you, one is a small engine that is deliberately modest, one ships a research-licensed model for quality, and one is a service with a UI.
Scope is worth stating. This page covers four tools that run on hardware you control. It does not cover a hosted voice service you can only subscribe to, and it does not cover a voice-cloning model that cannot be run locally, because neither changes where the audio is computed or who holds the clone.
Library, engine, service: three shapes
The four split into three shapes, and the shape predicts almost everything about adoption.
A library. coqui-ai/TTS is a Python library for advanced text-to-speech generation, with pretrained models in more than a thousand languages, tools for training new models and fine-tuning existing ones in any language, dataset analysis and curation utilities, a speaker encoder, several vocoder families, multi-speaker support, a Trainer API and detailed training logs on the terminal and in Tensorboard. Its model list covers spectrogram architectures, end-to-end models, and attention methods. It also carries a detail worth knowing: the README says the models marked with an asterisk in its performance section are internal and not released open source.
An engine. OHF-Voice/piper1-gpl describes itself as a fast and local neural text-to-speech engine that embeds espeak-ng for phonemization. The pieces are a CLI, an HTTP server, a Python API, a C and C++ library, documentation for training new voices, and a published voice list. That shape is why it turns up inside other projects: the README names Home Assistant, NVDA and LocalAI among the projects using it, and a Java binding exists as a separate project.
A service. devnen/Chatterbox-TTS-Server wraps Resemble AI's Chatterbox family, described as Original plus Multilingual plus a 350M-parameter Turbo model, behind an OpenAI-compatible API and a web UI, with 23 languages, voice cloning, chunking for large text, audiobook generation, built-in voices and a seed for reproducible output. fishaudio/fish-speech is also a service, with command line inference, a web UI, a server mode and a Docker setup, built around a 4B Dual-Autoregressive model.
Solo, small team, compliance, embedding
Solo. Piper1 GPL is the least ceremony: it installs from pip and the CLI produces audio on the machine in front of you. If you would rather not install anything at all, the Chatterbox README points to a Google Colab notebook for trying the server with voice cloning and audiobook generation.
Small team. If the group wants a shared service, Chatterbox-TTS-Server is the one shaped for it: a web UI, a configurable port, OpenAI-compatible endpoints, predefined voices and audiobooks. If the group wants a library inside an existing Python service, Coqui TTS is the mature choice, and our analysis of Coqui TTS describes it as suited to teams that want to synthesize speech, clone a voice from a short reference clip and fine-tune on their own dataset without leaving the Python ecosystem.
Company with compliance needs. Three separate questions, in order. Where does the audio go: nowhere, in all four cases, once the model is local. What may you do with the code: Piper1 GPL is GPL-3.0 and our analysis says to skip it if you cannot accept copyleft, Coqui TTS is MPL-2.0, Chatterbox-TTS-Server is MIT, and Fish Speech's research licence needs review before commercial deployment. And what rights do you have over the voice: that analysis of Chatterbox-TTS-Server tells you to check the voice cloning consent requirements in the upstream model licence, which is a question about the speaker and not about the software.
Developer embedding it. Piper1 GPL has the widest interface surface of the four, with an HTTP server, a Python API and a C and C++ library behind the same engine. Chatterbox-TTS-Server speaks the OpenAI audio API shape, so existing client code is the integration. Fish Speech is the one to look at if you serve at scale rather than per request: its README points at an SGLang server and a vLLM Omni recipe.
What self-hosting each one costs
The install commands describe four quite different amounts of work.
Coqui TTS: one pip install into an environment you control. The analysis tells you to verify that your Python version falls inside the range the setup script enforces, that the GPU you plan to use is covered by the CUDA image the Dockerfile builds on, and that the model you intend to use is listed under Released Models rather than Experimental Models, which is the difference between a documented model and an experiment.
Piper1 GPL: one pip install, then a voice file. The operational cost is close to zero, and the checks are the ones that analysis names: confirm your language is in the VOICES documentation, run the CLI once on the hardware you will deploy to, and read the CLI and Python API docs to decide which interface your application should use.
Chatterbox-TTS-Server: a requirements install plus pinned extra packages, including an onnx build and a tokenizer, and Python 3.10 on the host. The container route is the more interesting one, because the README ships a compose file per GPU stack, with CUDA 13.0 and PyTorch 2.10 for DGX Spark, cu121 and cu128 for RTX 30, 40 and 50 cards, ROCm 7.2 for AMD Strix Halo, plus Apple Silicon and CPU fallback. That analysis also tells you to confirm the chatterbox-tts package installs cleanly against your torch build.
Fish Speech: the largest footprint. Our analysis records no install command for it, and the README's routes are the official installation documents, a Docker setup, or the SGLang and vLLM Omni server recipes. The cheap way to evaluate it is the live demo at fish.audio and the Gradio web UI the analysis places on port 7860, before you commit any hardware.
The licence question, and the voice question under it
The licences are worth sorting out before quality, because two of the four have terms that a company has to look at properly.
Chatterbox-TTS-Server is MIT, which is the simplest case for the software. Coqui TTS is MPL-2.0, a file-level copyleft: the licence is permissive for use, and the obligation attaches to files you modify and distribute. Piper1 GPL is GPL-3.0, the strongest copyleft of the four, and our analysis of Piper1 GPL says to adopt it only if you can live with those terms. Fish Speech is the outlier: the README puts both the codebase and the model weights under the FISH AUDIO RESEARCH LICENSE and says violations will be acted on, and our analysis of Fish Speech is that this is not a standard open-source licence and requires review before any commercial deployment.
Underneath all four is a question the licences do not answer. Cloning a voice is a legal act, not a technical one, and the software licences here are silent on whether the person whose voice you cloned agreed. Our analysis of Chatterbox-TTS-Server points at the upstream model licence's consent requirements, and that is the right place to look for every candidate, including the ones whose code is permissively licensed.
There is also a model-licence layer to check separately from the code licence. Fish Speech is the clearest case where the weights are explicitly named alongside the code; for the others, the model files and their terms are a separate question from the repository.
Bottom line
A team that needs an HTTP endpoint, a web UI and permissively licensed code should start with Chatterbox-TTS-Server, accepting that it is a single-maintainer project and that the voice cloning consent rules come from the model, not the server. A Python team building speech into an existing service should use Coqui TTS and plan for a codebase whose last push was 2024-08-16. Piper1 GPL is the right answer for offline, low-latency, accessibility-oriented output where GPL-3.0 is acceptable and expressiveness is not the goal. Fish Speech is the one to evaluate when multilingual prosody matters, and to put past legal before infrastructure. Before committing, check the voice rights, then the model licence, then the code licence, in that order.