Clone a voice on your own PC

Record ten seconds of someone talking, type a script, and Mimic says it in their voice. Free and open source. Nothing leaves your computer.

Download for Windows Version 1.4.2 · Windows 10/11 and Linux
Typing a line in Mimic, generating it in a cloned voice with Qwen3-TTS, and playing the take back, recorded at real speed

One window, start to finish

A real recording at real speed: Qwen3-TTS 0.6B cloning a demo voice on an RTX 5070 Ti. Voices on the left, your script in the middle, takes on the right.

Download Mimic Read the source

Built for the second take

Cloning is quick. The work is in trying lines, engines and settings until a take sounds right, so that's what the window is made for.

Nine engines, one windowSwitching unloads the last one first, so your GPU stays free.
Keyboard firstGenerate, play with Space, save with Ctrl S.
Transcripts write themselvesSpeech-to-text runs on your machine too.
Compare enginesSame line, every engine that fits. Blind if you like.
Nothing saved behind your backTakes wait in a cache until you keep one. WAV, FLAC or MP3.
Voice profilesRecord by reading a short passage, import a clip, or drop audio on the window. Each profile keeps its clip, transcript and favorite settings.
Real memory, not a guessThe bar at the top shows what your GPU or RAM is really using.
Use it from your phoneYour PC does the work. Home Wi-Fi or Tailscale. Off until you start it.
Mimic on Linux in Caelestia colors
At home on LinuxWayland and X11. Follows Caelestia's live colors.
1.95 s

to make one second of speech with Qwen3-TTS 0.6B on an RTX 5070 Ti. On a 16-thread CPU, 4.7 s.

Measured on a real PC

2.4 GB

of VRAM runs Qwen3-TTS 0.6B. No GPU? 5.6 GB of RAM.

0

accounts, trackers or uploads. Your voices stay on your disk.

What will your PC run?

Set your graphics card and its memory. Every figure here was measured with Mimic while cloning, unless it says otherwise.

Seven engines run on your GPU.

  • Qwen3-TTS 0.6B

    2.4 GB VRAM · 5.6 GB RAM · clip and transcript · Apache 2.0

    GPU or CPU

  • Chatterbox Turbo

    3.4 GB VRAM · 6.0 GB RAM · clip only, over 5 s · MIT

    GPU or CPU

  • Audio8 TTS Preview 0.6B

    3.6 GB VRAM · 6.1 GB RAM · clip and transcript · Apache 2.0

    GPU or CPU

  • Qwen3-TTS 1.7B

    4.7 GB VRAM · 10.9 GB RAM · clip and transcript · Apache 2.0

    GPU or CPU

  • VoxCPM2

    6.9 GB VRAM · clip and transcript · Apache 2.0

    NVIDIA GPU

  • Breeze TTS 2

    About 8 GB VRAM (published) · set up by hand · non-commercial

    NVIDIA GPU

  • KittenTTS 2

    8.8 GB VRAM · 14.5 GB RAM · clip only · Stellon Labs Community

    GPU, CPU possible

  • Fish S2 Pro

    About 15 GB VRAM (published) · Windows through WSL · set up by hand

    NVIDIA GPU

  • NeuTTS Air

    About 1 GB RAM (published) · Linux · clip of 3 to 15 s and transcript · Apache 2.0

    CPU, Linux

Measured as peak memory while cloning, on an RTX 5070 Ti and a 16-thread CPU. AMD Radeon (ROCm, Linux) and Intel Arc support is experimental. Each engine keeps its own license, which applies to what you make with it.

First, read ten seconds aloud

Type your name and read the passage on screen. Mimic already knows the words, so every engine can use the clip straight away. Or drop in a 3 to 30 second clip of one speaker.

RecordTranscribeWriteGenerate

Fair questions

Is it really free?

Yes. Mimic is MIT-licensed and has no accounts, plans or limits. The engines it downloads keep their own licenses: most allow commercial use, but Breeze TTS 2 and Fish S2 Pro are non-commercial, and KittenTTS 2 has conditions for businesses.

What do I need to run it?

Windows 10 or 11, or Linux. An NVIDIA card with 3 GB of memory runs the smaller engines comfortably, and with no GPU Qwen3-TTS runs on the CPU in about 6 GB of RAM. You also need uv to download engines and ffmpeg for audio, both one command to install.

Does anything leave my computer?

Your recordings, voices and takes never do. Mimic goes online only to check GitHub for updates and to download the engines you choose. If you turn on sharing, your other devices connect over your home network or Tailscale.

Can I clone anyone's voice?

Only with their permission. Your own voice is fine; for anyone else, ask first. Using a cloned voice to impersonate, deceive or defraud someone is illegal in many places. Mimic asks you to confirm this before your first clone, and marks saved takes as AI-generated.

Is there a Mac version?

Not yet. Mimic runs on Windows and Linux. Its sibling app Latent, which does the same for images, already runs on Apple Silicon.

Which engine should I start with?

Qwen3-TTS 0.6B. It needs about 2.5 GB of VRAM, works on the CPU too, and is fast. Download speech-to-text from the same list so transcripts write themselves.

How do updates work?

Mimic checks GitHub when it starts and offers the new version. Arch users build from the PKGBUILD and update through makepkg.

Mimic's dark theme with the Narrator voice selected

Hear it in your own voice

Version 1.4.2. Install it, download Qwen3-TTS from Settings › Engines, and read the passage.

Mimic in the light theme