Voicebox: a free local voice studio, and the clone is the weakest part
Kokoro's stock voices are fast and natural, a voice designed from one sentence is the surprise, and the cloned voice is the let-down. £0, no account, all of it on a CPU.
# Voicebox: a free local voice studio, and the clone is the weakest part
Kokoro's stock voices are fast and natural, a voice designed from one sentence is the surprise, and the cloned voice is the let-down. £0, no account, all of it on a CPU.
What the posts claimed
Two bookmarks made the same claim.
- Voicebox voice cloning on PC — clone a voice from a few seconds of audio.
- Voicebox open source voice studio — an open source voice studio, 33k stars.
Underneath: a free local alternative to ElevenLabs and WisprFlow. Clone a voice, generate speech in 23 languages, build multi-speaker "stories" with effects, and dictate into any app from a hotkey.
The alternative-to-ElevenLabs part is the honest part. The cloning part is the headline, and it turned out to be the weakest.
What it really is
A Tauri desktop app for macOS, Windows and Linux, wrapped around a FastAPI server with a React web UI. The server runs on its own in Docker, which is how we tested it.
Engines in v0.6.0: Qwen3-TTS for cloning, Qwen CustomVoice (9 presets), Qwen VoiceDesign (a voice built from a text description), LuxTTS (fast cloning on CPU), Chatterbox and Chatterbox Turbo, HumeAI TADA, OmniVoice, and Kokoro with 50 presets at 82M parameters. Whisper transcribes. There is an OpenAI-compatible /v1/audio/speech endpoint and an MCP server, so an agent can speak without a second integration.
There is also "Voicebox Cloud", and it is off unless you explicitly enable it. We found no telemetry in the backend, and we set HF_HUB_DISABLE_TELEMETRY=1 as well.
What we set up
Upstream's own three-stage Dockerfile at commit 8af7efe, with one change: CPU-only torch. Upstream pulls the CUDA build, about 5 GB of NVIDIA libraries we cannot use. Image bookmark-lab/voicebox:003, container voicebox-003 on 6 CPUs with a 12 GB cap and two named volumes — one for data, one for the model cache, which reached about 6 GB after the demo.
We read the install scripts before running them: Dockerfile, entrypoint, requirements. No piping a download into a shell. No account was created and no key was given to the app; the one text we generated was written by our own gateway, and that key never left our script.
What we measured
Four clips, every voice synthetic. No real person's voice is used anywhere in this.
| Clip | What it is | Time on a CPU |
|---|---|---|
| 1. Two stock voices (30 s) | a kids' science dialogue written by an LLM, read by Kokoro's af_heart and bm_george and laid out on a Voicebox Story timeline | about 3 s per line |
| 2. The "Radio" effect | one Kokoro line through the built-in effect preset | seconds |
| 3. A designed voice (8 s) | "a fictional elderly lighthouse keeper, warm, slow, slightly gravelly", made by Qwen VoiceDesign from that one sentence | 278 s cold (3.6 GB download), 94 s warm |
| 4. A clone of clip 3 (4.5 s) | zero-shot clone of the synthetic voice, saying new words, with LuxTTS | 66 s cold, 12 s warm |
Every clip was transcribed back with Voicebox's own Whisper and says what it should — Whisper hears "Mira" as "Myra" and nothing else is wrong. Eight unit tests for our helper code pass.
What worked
- Free and local after the first model downloads. No account, no key, no per-character bill.
- Kokoro's stock voices are genuinely good. They are fast on CPU — real time or better — and they sound natural. For anyone who has used Piper, this is a different world.
- VoiceDesign is the standout. A believable new voice from one sentence, which means there is never a reason to clone a real person to get a new voice.
- The API is complete enough to automate end to end: profiles, generation, stories, effects, export.
What did not, or needs care
- The first transcription after Whisper downloaded failed with "Tensor on device meta". A container restart fixed it.
- The Qwen engines are slow on CPU, about 11x slower than real time. Fine for a short line, pointless for a paragraph. A GPU would change this.
- The dictation hotkey and system-audio capture are desktop-app features. We tested the server in Docker, so we have not tried them.
- We did not measure how close the clone is. That one is for your ears.
How it sounded
Our own listening notes, in order:
- Two stock voices: interesting.
- Radio effect: definitely sounded like a radio. Good.
- Designed voice, the lighthouse keeper: a really good voice.
- The clone: a weird echo, and a frequency band that sounded odd. The clone is the weakest of the four.
What we would use it for
- Narration for videos instead of a flat offline voice — Kokoro or a designed voice, through the OpenAI-compatible endpoint.
- Two-voice, podcast-style explainers from a script, entirely local.
- A voice for our own agents, through the MCP server.
The verdict
8 / 10. Worth keeping, and the most useful tool we have tested. The hype is about cloning, but the best parts are the stock Kokoro voices and VoiceDesign. Together they give a free "script in, narrated audio out" path with a voice worth listening to.
The rule we would follow: stick to Kokoro for long text, and use a designed voice — used directly, not cloned — for anything that needs character. If you want to clone, clone a voice you already own.
Cost and accounts
£0. No accounts. Voicebox is MIT. The models come from Hugging Face and carry their own licences — Kokoro and Qwen3-TTS are Apache-2.0 — so check the others before using them commercially.
Responsible use
Voicebox ships a RESPONSIBLE_USE.md, and the rule is the obvious one: cloning a real person's voice needs that person's consent. Everything in this test is synthetic.