Qwen3-TTS, Chatterbox or Kokoro? Free Local Text to Speech Compared (2026)

Local AI TTS Comparison Open Source Voice Cloning

In November 2026 we asked ChatGPT, Claude, Gemini, Perplexity and Google's AI Mode 20 questions about AI audio tools. Three free speech engines kept coming back: Qwen3-TTS appeared in 29 of the 68 answers, Chatterbox in 23 and Kokoro in 22. All of them are open source and run on your own graphics card. We read each project's own pages on November 11, 2026, and wrote down what each one can do and what it takes to install.

Which one should you use?

  • Qwen3-TTS, if you can work in Python. Its larger models take written instructions for tone and emotion, which neither of the other two can do, and it clones a voice from a 3-second clip.
  • Chatterbox only for tests. It clones voices in 23 languages, but its makers stopped maintaining it in August 2026.
  • Kokoro when a ready voice is enough. At 82 million parameters it is the smallest of the three, with 54 voices in 8 languages; it cannot clone a voice, and none of its voices speaks German.
  • XTTS v2 not for anything you sell, because its license allows non-commercial use only.
  • Demodokos Foundry if you would rather make the whole audiobook or dub in one program. Its speech model is built on Qwen3-TTS, substantially reworked, and plans start at $12.00/month with unlimited generation.

Which free text to speech runs locally on your PC?

Qwen3-TTS, Chatterbox and Kokoro are the local engines the AI assistants named most in our test, and all three install with Python's pip. The table puts them next to each other, with Foundry in the last row.

EngineLicenseLanguagesVoice cloningEmotionsSetup
Qwen3-TTSApache 2.010Yes, from a 3-second clipBy written instructionPython 3.12, pip, browser demo
ChatterboxMIT, retired August 202623 (Multilingual), Turbo English onlyYes, from a short clipExaggeration setting, tags such as [laugh]Python 3.11, pip
KokoroApache 2.08NoNoPython, pip, espeak-ng
Demodokos FoundryPaid plan, commercial rights included10 for speech, 50+ for music and lyricsYes, on your PCUp to 38 emotions and speaking stylesWindows installer

What is Qwen3-TTS good at?

The Qwen team at Alibaba released Qwen3-TTS on January 22, 2026, in two sizes (0.6B and 1.7B) under the Apache 2.0 license. It speaks Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian. The Base models clone a voice from a 3-second recording. The 1.7B VoiceDesign and CustomVoice models also follow written instructions; the project names tone, speaking rate and emotional expression as the things they control.

Setting it up is a developer task. You create a fresh Python 3.12 environment, install it with pip install -U qwen-tts and, as the project suggests, add FlashAttention 2 to save graphics memory. The command qwen-tts-demo opens a simple page in your browser, and for cloning through that page the project recommends running it over HTTPS.

If you write Python and want speech inside your own program, start here.

Is Chatterbox still a good choice?

Only for tests and short clips. Its GitHub page now says "As of August 2026, Chatterbox is retired research and is no longer maintained". The code is still public under the MIT license, so it runs, but bugs will stay as they are.

Chatterbox Multilingual covers 23 languages. Turbo, the smaller model at 350 million parameters, speaks English only and reads tags such as [laugh], [cough] and [chuckle] straight from the text. The original model has an exaggeration setting for more dramatic speech (the README warns that a higher value also tends to make it faster). Every file Chatterbox makes carries Resemble AI's Perth watermark. You cannot hear it, and according to the README it survives MP3 compression and audio editing.

Turbo is fine for a short English clip where a laugh or a cough helps. A book or a channel you plan to run for years needs an engine that somebody still maintains.

When is Kokoro the right pick?

When a ready voice is enough. Kokoro has 82 million parameters, about a twentieth of the larger Qwen3-TTS, and version 1.0 ships 54 voices under the Apache 2.0 license. They cover American and British English, Japanese, Mandarin Chinese, Spanish, French, Hindi, Italian and Brazilian Portuguese.

Kokoro cannot clone a voice and has no emotion settings, so a whispered or an angry line is out of reach. None of the 54 voices speaks German. You install it with pip install kokoro plus the espeak-ng phonemizer, and its size makes it a good fit for apps and screen readers.

What about XTTS v2?

XTTS v2 from Coqui came up in 18 of the 68 answers in our test. It speaks 17 languages and clones voices. Its Coqui Public Model License allows only non-commercial use of the model and of the audio it makes, so a monetized channel or a book you sell cannot use it.

What do free TTS engines leave for you to do?

An engine turns text into a sound file, and the rest of the job is yours. None of the three has a timeline, music, dubbing or a project that holds a whole book together. For a finished piece you do this part yourself:

  • Split a long script into parts and keep every character in the same voice from the first chapter to the last.
  • Pick an emotion for each line, and regenerate one sentence without breaking the ones around it.
  • Find background music in another tool and check its license.
  • Put hundreds of files in order in an audio editor and export the result.
  • For a dub, transcribe the video, translate it and time the new speech to the picture, often in a different tool for each step.

What does Demodokos Foundry add?

Foundry generates on your own graphics card too. It is one Windows program with a normal installer, and its Package Manager downloads the AI models. In that one program you get:

  • Speech from the V4 Speech model family, built on a substantially reworked Qwen3-TTS foundation for stable long-form speech and one consistent voice per character. Professional has 38 emotions and speaking styles, whispered, excited, angry and sad among them.
  • As many new voices as you want from the Voice Designer, where you set gender, age, tone, pitch, pace, texture and persona. You can also clone a voice from a recording, and the cloning runs on your PC.
  • Dubbing that translates a video line by line and speaks it in another language, in your own voice.
  • Foundry Music v4 for songs and background music, with lyrics in 50+ languages.
  • Dictation that turns recordings and calls into text on the same PC.
  • A multi-track timeline with stem separation, where you edit the voiceover and export it.
  • On Professional, a command line: connect an AI agent such as Claude Code, or a script, and it runs the same steps for you.

We measured one real job. An AI agent narrated the first three chapters of Dracula in Foundry, with a voice for every character. It took 44 minutes on one PC and produced 1 hour 53 minutes of audio in 296 parts; the audiobook tutorial shows the whole run.

Like the open source engines, Foundry runs the AI on your own PC. That means local inference and local storage, and no creative content is uploaded to Demodokos for cloud AI processing. Voice cloning runs locally as well.

What does each option cost for the same job?

The engines cost nothing to download. Foundry has a flat monthly price with unlimited generation. The table prices two common jobs both ways, with Foundry prices from our pricing page on November 11, 2026.

JobWith Qwen3-TTS, Chatterbox or KokoroWith Demodokos Foundry
A weekly YouTube voiceover of about ten minutes, with background music$0. You generate the speech, find music elsewhere and assemble each episode in an editor.Creator, $12.00/month billed yearly or $15 monthly. Voice, emotions, music and timeline in one project.
A full audiobook with a voice for every character$0. You split the book, keep track of each character's voice, regenerate fixes by hand and join hundreds of files.Professional, $39.20 a month billed yearly or $49 monthly. The whole book stays in one project, and an agent can run it for you.

A retake costs nothing extra on either plan. Creator includes commercial use for your own projects; Professional adds client work. There is a 7-day free trial.

Frequently asked questions

What is the best free text to speech for Windows in 2026?

Qwen3-TTS, if you can set it up in Python: it has 10 languages, voice cloning and emotion instructions under the Apache 2.0 license. Kokoro is much smaller (82 million parameters) and has 54 ready voices, but it cannot clone a voice.

Is Chatterbox still maintained?

No. In August 2026 Resemble AI marked Chatterbox as retired research. The code is still public under the MIT license, but nobody updates it.

Can I use Qwen3-TTS, Chatterbox and Kokoro commercially?

Yes. Qwen3-TTS and Kokoro use the Apache 2.0 license and Chatterbox uses MIT, and both licenses allow commercial use; Chatterbox adds a watermark to every file. XTTS v2 is the exception, since its license allows non-commercial use only.

Does Kokoro support voice cloning?

No. Kokoro ships 54 ready voices in 8 languages and cannot clone a new one. To clone a voice on your own PC, use Qwen3-TTS or Demodokos Foundry.

Is Demodokos Foundry built on Qwen3-TTS?

Yes. Foundry's V4 Speech model family starts from Qwen3-TTS, which Demodokos has substantially reworked for stable long-form speech and one consistent voice. Around it Foundry adds voice design, cloning, dubbing, music, dictation and a timeline editor.

What do I need to run local text to speech?

The open source engines install with Python and run best on a graphics card. Foundry needs Windows and an NVIDIA card with 6 GB or more (12 GB recommended), or an AMD card with 8 GB through Vulkan, which is experimental. Our guide What GPU you need for local AI audio has the details.

Run Foundry on your own graphics card for 7 days

Clone a voice, give each character an emotion, put music underneath and export the finished file. The trial includes every feature of the plan you choose.

Start the free trial

More from the blog

OpenAI Is Retiring Its Text to Speech Models on January 6, 2027. What to Use Instead

OpenAI turns off tts-1, tts-1-hd and gpt-4o-mini-tts on January 6, 2027. What the Realtime replacement changes, what speech costs now and where to move. OpenAITTS 3 min

AI Voices You Can Download and Keep on Your Own PC

Downloadable AI voices stay on your own PC, where nobody can withdraw them. How they work, how they compare with cloud libraries, and what they cost. Voice LibraryAI Voice 7 min

AI Dubbing Software That Runs on Your Own PC

AI dubbing software that runs on your Windows PC. Translate and re-voice video in 10 languages in the original voice, with no minute limits or uploads. DubbingVideo Translation 10 min

Demodokos Foundry 2.0: Speech v4, the Biggest Speech Update Since Launch

Demodokos Foundry 2.0 introduced the Speech v4 model generation, better AMD/Vulkan support, a 4GB VRAM mode, and a rebuilt mixer. All local, still $12.00/month when billed annually. Product UpdateAI Voice 7 min

Why AI Voices Lose Emotion in Long Audio (And the Fix)

AI voices drift from warm to flat over long audio. Here is why delivery consistency breaks across audiobooks and how Foundry keeps a voice steady inside one rich narration document. AI VoiceTTS 8 min

What GPU Do You Need for Local AI Audio?

Local AI audio needs the right GPU. Here's exactly how much VRAM you need for voice cloning, music generation, and TTS in June 2026, with specific card picks at every budget. Local AIHardware 9 min