A scope note first: this site has already covered VoiceStudio as a single product and mapped the cloud-versus-local decision. This article does neither. It puts five local voice cloning toolkits on one table and picks between them by what you are actually trying to do — which project turned cloning into what shape, and who should grab which tool.
The five could hardly differ more. VoiceStudio is a ready-to-run desktop app. GPT-SoVITS is a training framework. fish-speech follows a model-plus-service route. IndexTTS is industrial zero-shot TTS. CosyVoice, from Alibaba's FunAudioLLM, is the control group: the widest license attached to a repo that has stopped moving. Star counts via the GitHub API on 2026-10-15: VoiceStudio 66,399, GPT-SoVITS 62,825, fish-speech 33,132, IndexTTS 24,481, CosyVoice 23,972. Stars are not quality, but they sketch the attention hierarchy — the two leaders have left the rest at less than half their size.
Compare the Shape, Not the Sound
One discipline upfront: without a common test set and playback environment, subjective quality scores tell you nothing, so this review contains none. The real differences are structural — shape decides how hard a tool is to start, what workflows it plugs into, and how carefully you must read its license. Every cell below has a source: stars and status from the GitHub API snapshot of 2026-10-15, every license checked against the repository's LICENSE file.
<table> <thead> <tr><th>Project</th><th>Stars (2026-10-15)</th><th>License (verified)</th><th>Form</th><th>Entry barrier</th><th>Best fit</th></tr> </thead> <tbody> <tr><td>VoiceStudio</td><td>66,399</td><td>AGPL-3.0 (LICENSE reviewed)</td><td>Electron desktop app, out of the box</td><td>Lowest: install and go</td><td>Creators who need results today</td></tr> <tr><td>GPT-SoVITS</td><td>62,825</td><td>MIT (LICENSE reviewed)</td><td>Training framework</td><td>Medium: run the training yourself</td><td>Your own voice as a model</td></tr> <tr><td>fish-speech</td><td>33,132</td><td>Fish Audio Community License v1.0 (reviewed, custom)</td><td>Model + service</td><td>Medium: local or hosted</td><td>API products (mind the thresholds)</td></tr> <tr><td>IndexTTS</td><td>24,481</td><td>bilibili Model Use License (reviewed, custom)</td><td>Industrial zero-shot TTS</td><td>Medium: no fine-tuning needed</td><td>Engineering zero-shot scenarios</td></tr> <tr><td>CosyVoice</td><td>23,972</td><td>Apache-2.0 (LICENSE reviewed)</td><td>Model + training</td><td>Medium: developer-oriented</td><td>Wide license (mind the stalled repo)</td></tr> </tbody> </table>One line before the details: use it now, pick VoiceStudio; train your own voice, GPT-SoVITS; API for a product, fish-speech — after reading its license; industrial zero-shot, IndexTTS — same advice; Apache and total freedom, CosyVoice — while accepting five months without a single commit.
VoiceStudio: Cloning in a Desktop App
VoiceStudio is the only one where "install and it works" is literally true: an Electron desktop app with no command line and no self-assembled inference stack. The default engine is k2-fsa/OmniVoice (repo: github.com/k2-fsa/OmniVoice, 14,570 stars, Apache-2.0), officially supporting 646 languages.
Version 0.5.6 pushed the boundary further. A Twilio integration lets a cloned voice answer phone calls — off by default, verified via signed webhooks, single-use stream tokens, so the anti-abuse machinery ships with the feature. The same release added integrations with Claude Code, Cursor, Codex CLI, and the OpenAI Agents SDK, long-form audiobook generation on an 8GB GPU, and swappable voice samples. Developers get a local OpenAI-compatible API on port 3900 plus MCP support — click through the GUI, or let agents call it.
The license is AGPL-3.0, verified against the LICENSE file. Strong copyleft: personal use is unaffected, but distributing a modified version — especially as a network service — obligates you to open the derivative code under the same license. Creators will never feel this line; closed-source product teams must think before they type.
Repo: debpalash/VoiceStudio, also at github.com/debpalash/VoiceStudio.
GPT-SoVITS: One Minute of Audio, Your Voice as a Model
GPT-SoVITS takes the other road: instead of handing you a pleasant stock voice, it turns your own recordings into a model. Per the official README, fine-tuning works with about one minute of data — a real selling point when recording time is the scarce resource. A single clean self-introduction clip gets you started.
It is the veteran here and arguably the loudest voice-cloning option in the Chinese open-source community; 62,825 stars (GitHub API snapshot 2026-10-15) mean tutorials, dataset recipes, and pitfall answers are easy to find. The barrier is real — you process the dataset, launch the training, deploy the inference — but you buy full control over the voice and the process.
The license is MIT, verified against the LICENSE file: commercial use, modification, and redistribution carry no gate beyond keeping the copyright notice. For training your own voice and shipping it in a product, this is the route with the least licensing friction of the five.
Repo: RVC-Boss/GPT-SoVITS, also at github.com/RVC-Boss/GPT-SoVITS.
fish-speech: SOTA Reputation, a License Worth Reading
fish-speech is commonly placed in the SOTA tier of open-source TTS, and its model-plus-service shape means local deployment or a vendor-hosted offering. 33,132 stars (GitHub API snapshot 2026-10-15) and an active commit stream show it is still moving.
Before choosing it, one document is mandatory: the Fish Audio Community License v1.0, updated October 13, 2026 — every clause checked here against the LICENSE file. It has three layers. Research and non-commercial use are free. Commercial use is permitted while annual revenue stays under USD 500,000 (the "Revenue Threshold"); at or above it, you must notify Fish Audio in writing within thirty days, a ninety-day transition period follows, and continued use beyond that requires a separately negotiated written license. Hosted services are walled off entirely: regardless of revenue, making the model available to third parties so they can generate speech from their own inputs — via API, SDK, model hosting, or a site or app whose primary purpose is TTS, voiceover, or cloning — needs a separate written license. Explicit non-hosted exceptions: speech for content you control inside your own product; production services delivering only finished output to clients without self-serve access; free non-commercial research demos such as a Hugging Face Space.
Easy-to-miss red lines: pass the agreement along when distributing and display "Built with Fish Audio" prominently; no cloning an identifiable person's voice without consent; no stripping watermarks or provenance signals; no using outputs to train other speech models. Outputs you generate belong to you, to the extent permitted by law.
The short version: a small team wiring an API into its own product is explicitly permitted; a mass-market cloning platform is not, at any revenue, without a separate deal. This is a terms summary, not legal advice.
Repo: fishaudio/fish-speech, also at github.com/fishaudio/fish-speech.
IndexTTS: Bilibili's Industrial Zero-Shot Play
IndexTTS comes from Bilibili and targets industrial zero-shot TTS: a few seconds of reference audio in, a cloned voice out, no training loop. 24,481 stars (GitHub API snapshot 2026-10-15); the latest commit landed 2026-09-29, so maintenance is alive.
The custom bilibili Model Use License, checked here against the LICENSE file, does not ban commercial use but draws two hard lines. First, a scale gate: if your or your affiliates' products topped 100 million monthly active users in the prior month, or RMB 1 billion in annual revenue in the prior year, you must request a separate license that Bilibili may grant at its sole discretion — until granted in writing, you are not authorized. Second, model isolation: you may not use bilibili indextts2 or derivatives to improve any other AI model, except IndexTTS itself, its derivatives, or non-commercial models. Downstream recipients must be bound by the same terms, notices retained, and derivative distributions must state that modifications are not endorsed by the original right-holder.
Also recorded: high-risk deployments (medical diagnosis, autonomous driving, military, critical infrastructure, large-scale biometric surveillance, automated decisions) are prohibited or left to you to satisfy every compliance obligation. The agreement is governed by Chinese law, disputes go to the Shanghai Arbitration Commission, and the Chinese text prevails over the English — overseas users are effectively signing under Chinese law.
Repo: index-tts/index-tts, also at github.com/index-tts/index-tts.
CosyVoice: The Widest License on a Frozen Repo
CosyVoice comes from Alibaba's FunAudioLLM and is the only Apache-2.0 project of the five, verified against the LICENSE file: commercial use, modification, redistribution, derivative models — no added conditions. The widest license here, on a 23,972-star community (GitHub API snapshot 2026-10-15).
Its control-group value is the reminder it delivers: the last code push was 2026-05-25, nearly five months of silence, and the repository has moved — FunAudioLLM/CosyVoice now redirects to QwenAudio/CosyVoice, stars and history intact. Apache-2.0 means you can run it, fork it, and maintain it yourself; a permissive license plus a frozen upstream is a common combination. But if "the community keeps patching bugs" is part of your criteria, weigh those five months. The listening half of voice AI was covered in a separate streaming ASR comparison; this piece is strictly about speaking.
Repo: FunAudioLLM/CosyVoice (redirecting to QwenAudio/CosyVoice), also at github.com/QwenAudio/CosyVoice.
The Decision Route
Licenses and shapes together make the routes clear.
Need it today: VoiceStudio — works out of the box, v0.5.6 already ships phone answering and agent integrations, and AGPL-3.0 burdens personal use not at all.
Your own voice as a model: GPT-SoVITS — a minute of data per the official README, the largest Chinese-community knowledge pool, MIT license.
An API for your product: fish-speech, provided revenue stays under USD 500,000 and the use is not a hosted service; delivering finished audio or voicing your own content is explicitly covered. Beyond that, contact Fish Audio first.
Industrial zero-shot without a training loop: IndexTTS — seconds of reference audio, engineering-ready; check the two hard lines before crossing 100 million MAU, RMB 1 billion revenue, or improving another model.
Apache and total autonomy: CosyVoice — zero license friction, priced in five months of upstream silence; choosing it means being your own maintainer.
One honest caveat before you pick: every star count and license clause in this piece was verified against the GitHub API and the LICENSE files on 2026-10-15 - snapshots, not eternal truths. Repos drift, licenses get rewritten, and a community license can change its thresholds the day after you build on it. Before any commercial commitment, re-read the LICENSE file of the week, not the blog post you remember. And if you take only one thing from this comparison, take this: the tool that survives on your machine is the one whose maintenance model you can live with - an active repo with a strict license, or a frozen repo with a free one, is a real trade-off, not a ranking.
Related reading
- Which Open-Source Streaming ASR Never Rewrites Your Words?
- VoiceStudio: the local-first open-source voice studio
- Cloud vs Local Voice AI: A Cost and Control Showdown
FAQ
Q1: Do all five run fully locally?
A1: Yes by positioning: VoiceStudio is a local desktop app that also opens a local API port, and GPT-SoVITS, fish-speech, IndexTTS, and CosyVoice all support local inference. fish-speech coexists with a hosted service, but this review covers only its local form. Hardware requirements live in each repository's docs; we did not benchmark them.
Q2: Can fish-speech be used commercially?
A2: Yes, under the Fish Audio Community License v1.0, for entities with annual revenue under USD 500,000. Reaching the threshold triggers a thirty-day written notice and a ninety-day transition, after which a separately signed license is required. Hosted services need a separate written license at any revenue. Voicing your own content, delivering finished audio, and free research demos are explicitly exempt.
Q3: Who does the IndexTTS license constrain most?
A3: Very large users — affiliates above 100 million monthly active users or RMB 1 billion annual revenue need a separate license at Bilibili's sole discretion — and teams wanting to reuse its capabilities to improve other AI models, which is barred except for IndexTTS itself, its derivatives, or non-commercial models. Chinese law governs and the Chinese text prevails.
Q4: I just want to clone my own voice for my videos. Easiest pick?
A4: VoiceStudio. It works right after installation, voice samples are swappable, and AGPL-3.0 poses no practical burden on personal use. Turn to GPT-SoVITS's training route only for deeper customization.
Q5: CosyVoice has stalled — still usable?
A5: Yes. Apache-2.0 stands, code and models remain available, and the repo simply went nearly five months without a commit while moving under the QwenAudio organization. Choosing it means accepting maintenance duty; if you can hold a fork or need a stable tool as-is, it remains the most permissive option.