free-voice-clone is a curated map of local text to speech models

One file, one table, dozens of models

The repository is a single README.md and nothing else, which is the right shape for what it is. The stated purpose is a list of all local and free open source voice-clone text to speech models and music generation models. There is no code, no build system and no configuration, so there is nothing to install and nothing to break. The value is the comparison table, and it is dense rather than decorative. Each row names a model and then answers the questions you would otherwise have to open twenty repositories to answer: does it clone a voice zero-shot, does it also do speech recognition, how many languages does it cover, can it stream, and what licence governs the weights. The table currently lists three dozen text to speech models including Voxtral-4B-TTS-2603, LongCat-AudioDiT, Fish Audio S2 Pro, KittenTTS, MOSS-TTS, SoulX-Singer, Qwen3-TTS, Kokoro-82M, IndexTTS2, Chatterbox, Fish Speech, Dia and ZipVoice. A few rows are marked as not supporting voice cloning, such as Supertonic 2, which tells you the table is a survey of the TTS field rather than a voice cloning list only.

Licence is a column, and it is the column that matters most

Reading the licence column across the table is the most useful thing this repository does, because licence terms in this field are a trap for anyone planning to ship. Most of the listed models are Apache-2.0, including MOSS-TTS, Qwen3-TTS, GLM-TTS, IndexTTS2, Dia, Fish Speech, ZipVoice, VoxCPM, LuxTTS, MegaTTS3, Spark-TTS, Step-Audio, Step-Audio-EditX, FireRedTTS2, Fun-CosyVoice 3.0, NeuTTS, SoproTTS, SoulX-Singer, SoulX-Podcast, VieNeu-TTS, MiMo-Audio and the Kokoro and KokoClone entries. Others are not. Fish Audio S2 Pro is listed under a Research License and in its own entry under the Fish Audio Research License. MioTTS-2.6B and LFM2-Audio-1.5B carry LFM terms, Supertonic 2 is under OpenRAIL-M, and Kimi-Audio is MIT/Apache-2.0. The entries also make the scale of the weights visible, since Fish Audio S2 Pro is noted at around 10 GB in BF16, while KittenTTS is described as a model under 25MB with 15 million parameters that runs without a GPU on any device. The same category of caveat applies to the project's own licensing, since the repository declares no licence of its own and its primary language is recorded as unknown. Treat the list as a lead generator, and read the licence of each model directly rather than trusting the shorthand in the table.

What the detail entries actually tell you

Each expandable entry follows the same template, and the template is where the useful detail lives. A description sentence, a release date, then a table of parameters, zero-shot voice cloning support, ASR support, pronunciation control, emotion control, languages, streaming, sample rate and licence, followed by links to the GitHub repository and the Hugging Face weights. The descriptions are where the technical differences surface. LongCat-AudioDiT is a diffusion-based model operating directly in waveform latent space from the Meituan LongCat team, needing only a Waveform VAE and a diffusion backbone, and its stated key innovation is Adaptive Projection Guidance replacing traditional classifier-free guidance. Fish Audio S2 Pro is described as combining reinforcement learning alignment with a dual-autoregressive architecture, with fine-grained inline control of prosody and emotion, listed at 5B parameters as 4B slow autoregressive plus 400M fast autoregressive, and streaming figures given as a real time factor of 0.195 and 100ms time to first audio. MOSS-TTS is described as a production-grade foundation model from the OpenMOSS team and MOSI.AI, with a delay variant at 8B and a local variant at 1.7B. These are the details that decide a project, and the fact that the list records them per model is what makes it worth reading rather than skimming.

The table covers more than speech synthesis

The table of contents is broader than the name of the repository suggests, and the sections are the map of the field. There is Text-to-Speech Models, Music Generation Models, Anything to Audio, Audio Restoration and Enhancement, Speech Recognition, and Additional Resources. A handful of the TTS entries also carry ASR support in the comparison table, which is a real differentiator rather than a footnote. Voxtral-4B-TTS-2603 is marked as supporting both voice cloning and ASR across 50 or more languages with streaming under Apache-2.0. LFM2-Audio-1.5B, Step-Audio, MiMo-Audio and Kimi-Audio also show ASR support. Speech recognition as a separate section matters for anyone building a voice agent, because the loop of transcribe, reason and speak needs a recogniser and a synthesiser that agree on the same audio format and latency budget. Streaming is the other column worth watching, since a non-streaming model that emits a whole file at the end is unusable for a live voice interface, and LongCat-AudioDiT is one of the entries marked as not streaming. The Anything to Audio and Audio Restoration sections are adjacent territory: the first covers general sound generation, the second covers cleaning up recordings, which is often the missing piece when you have model output you cannot play back cleanly.

What a list cannot do for you

This repository is a directory, and the limits of a directory are worth stating plainly. It cannot tell you how a model sounds. Voice similarity, intelligibility and naturalness are the things you actually need to know, and no table of parameters substitutes for listening to the same sentence rendered by the three candidates. It cannot tell you how a model behaves on your data, and voice cloning quality is sensitive to the reference audio you feed it. It cannot tell you whether a model will fit your hardware, since parameter counts translate to memory use in a way that depends on quantisation, runtime and the precision you run at. It also cannot tell you whether a model still works, and this is the practical problem with the format. The last push to the repository was on 2026-04-09, more than six months before the current date, and the project has no GitHub releases, so there is no version history to check and no changelog. The list was accurate when it was written, and weight repositories, licences and even project names move. Treat every row as a pointer to verify rather than a fact to rely on, and check that the linked GitHub repository and Hugging Face page still exist before you plan around an entry.

How to use the list without wasting a week

A practical path through the list starts with the licence column, because it removes candidates fastest. Filter to Apache-2.0 and MIT if you intend to ship, and see how much of the field survives; a substantial part of the table does, which is the useful finding. Next filter on streaming if you are building anything interactive, and then on language coverage, since the counts in the table are specific enough to act on. Fish Audio S2 Pro lists 80 or more languages with 23 EU languages for KugelAudio, 23 or more for Chatterbox, 50 or more for Voxtral-4B-TTS-2603, and narrower sets such as 10 for Qwen3-TTS, 8 for Kokoro-82M with 54 voices, and single-language entries for Dia, SoproTTS, Maya1, LuxTTS and IndexTTS2's Zh and En pair. Then filter on size, because the table gives you both parameter counts and, where stated, weight sizes. What is left should be two or three models, and at that point the table has done its job. The remaining work is reading the model cards for those two or three and trying them on your own reference audio, which is where the decision actually gets made. The list is the funnel, not the answer, and treating it as the answer is the mistake this kind of repository invites.

Editorial conclusion

Use free-voice-clone as a shortlist and a comparison sheet when you are choosing a text to speech model to run on your own hardware, because the per model entries record parameters, languages, streaming support, sample rate and licence in a consistent form. Do not treat it as a library you can install, because the repository contains only a documentation file, and do not copy a licence judgment from the table without reading the model card, since several entries carry research, LFM or OpenRAIL terms rather than Apache-2.0. Before you commit to one, verify the model weights still exist at the linked repository or Hugging Face page, check the hardware your machine has against the parameter count, and read the licence text for the specific model you picked.

Frequently asked questions

Official sources

1. 0xSojalSec/free-voice-clone on GitHub
2. Issues
3. README