Comment "voice" on a YouTube vlog right now and you'll get a link to Voicebox — pitched as a free, local ElevenLabs alternative that clones a voice from three seconds of audio. The pitch is mostly accurate. The three-second number, and the licensing, come with an asterisk.
Voicebox v0.5, April 2026: Dictation, Personality Presets, and the Whisper Layer
Voicebox (repo jamiepine/voicebox) is a free, MIT-licensed desktop app that bundles seven text-to-speech engines, system-wide Whisper dictation, and a local REST/MCP surface into one local-first voice studio. It is unrelated to Meta's same-named 2023 research model, which Meta announced June 16, 2023 but declined to release over misuse risk .
The app is built by Jamie Pine, author of the open-source file explorer Spacedrive. Its first public release, v0.1.0, shipped January 27, 2026, and the current v0.5.0 "Capture" release landed April 25, 2026 . The homepage claims 1M+ downloads . Architecturally, it is a React frontend in a Tauri shell over a Python FastAPI backend on port 17493 .
| Engine | Model size | Zero-shot cloning | Languages |
|---|---|---|---|
| Kokoro (82M) | ~350 MB | No (presets) | English |
| Qwen3-TTS 0.6B | ~1.2 GB | Yes | Multilingual |
| Qwen3-TTS 1.7B | ~3.5 GB | Yes | Multilingual |
| Chatterbox Multilingual | ~3.2 GB | Yes | 23 |
| Chatterbox Turbo | — | Yes | Multilingual |
| HumeAI TADA 3B | ~8 GB | Yes | Multilingual |
| LuxTTS | — | Yes | Multilingual |
One caveat before you build: the MIT license covers the wrapper app, not the bundled weights . Each upstream model — Alibaba's Qwen, Chatterbox, TADA, Kokoro — carries its own terms you must check separately.
Disk, RAM, and a Clean Microphone: Preparing Your Seed Recording
Before you clone anything, confirm your machine clears the bar. Voicebox lists a documented minimum of macOS 11+, Windows 10+, or Linux, plus 8 GB RAM, 5 GB storage, and a modern multi-core CPU; the recommended tier is 16 GB+ RAM, 10 GB+ storage, and a CUDA-capable NVIDIA GPU . GPU backends span NVIDIA (CUDA), AMD (ROCm), Intel Arc (DirectML), and a CPU fallback .
Budget time and disk for first launch: the chosen engine downloads its model automatically . Sizes range from Kokoro at ~350 MB to Chatterbox Multilingual ~3.2 GB and TADA 3B ~8 GB; Whisper tiers add Base ~300 MB, Small ~500 MB, Medium ~1.5 GB, Turbo ~1.5 GB, and Large ~3 GB . On Apple Silicon, MLX-Whisper is auto-preferred over PyTorch .
Linux users have no prebuilt binary yet, so run it via Docker : git clone the repo, docker compose up, then open http://localhost:17493. The container binds to 127.0.0.1:17493 and mounts persistent data plus a Hugging Face cache volume — but ships with no built-in auth beyond that localhost binding , so do not expose the port.
Cloning a Personality in Voicebox: the Numbered Procedure
Once Voicebox is running, cloning a voice is a six-step flow through the Profiles panel — no config files, no API keys. The steps below assume you already prepared a clean sample. Install and launch first, then work top to bottom.
- Install for your platform. Grab the macOS DMG (Apple Silicon or Intel), the Windows MSI, or run the container on Linux, which has no prebuilt binary yet .
- Launch and open the Profiles panel. This is where both cloned and preset voices live .
- Create a Cloned profile and pick a cloning-capable engine. Zero-shot cloning works with Qwen3-TTS, Chatterbox Multilingual, Chatterbox Turbo, LuxTTS, or TADA. Kokoro (82M) is preset-only and does no cloning .
- Upload or record your reference. The docs recommend a 10–30 second WAV of natural, complete sentences in a quiet room; WAV is preferred over MP3, M4A, or FLAC .
- Generate a test phrase. Verify tone and clarity before committing the profile — quality varies by engine and sample.
- Assign the personality to dictation or TTS output.
If you only want to compare engine behavior, skip the recording entirely: Voicebox ships 50+ preset voices that store no sample and point to built-in voices . Use one to hear how each engine renders speech, then record your own sample once you have picked the engine that sounds right.
What the Homepage Omits: Realistic Seed Expectations and TTS Obligation
The homepage's "a few seconds" is marketing shorthand, not the documented spec. Voicebox's own quick-start and profile docs recommend a longer, cleaner reference than the 3-second figure circulating in demos:
"10–30 seconds of clear speech with minimal background noise and natural, complete sentences," per the official profile documentation (source: Voicebox docs, 2026-04).
A 3-second clip may work with some zero-shot engines, but it sits below that range and amplifies sensitivity to noise, compression, and atypical delivery . For contrast, ElevenLabs Instant Voice Cloning recommends at least one minute — one to two ideal — on its Starter plan, and Professional Voice Cloning needs the Creator plan at $22/month plus roughly 30 minutes of audio for genuine fine-tuning, not runtime conditioning .
Two more items the download button glosses over. First, licensing: the MIT license covers the jamiepine/voicebox wrapper, not the bundled weights — Qwen, Chatterbox, TADA, and Kokoro each carry separate upstream terms you should verify before production use . Second, exposure: the backend binds to 127.0.0.1:17493 by default; remapping it to 0.0.0.0 removes the only barrier, because no credential system is built in . And on fidelity, no independent reproducible benchmark against ElevenLabs existed as of April 2026 — the "free local alternative" case rests on ownership and cost, not verified parity.
Keep consent explicit. The illustrative snippet below (not part of Voicebox, but it did execute) encodes the fine print as a hard gate:
"""Minimal Voicebox-style demo: 3 seconds in, but consent stays explicit."""
class Voicebox:
def clone(self, voice_sample_seconds, text, consent=False):
if voice_sample_seconds < 3:
raise ValueError("Need at least a 3-second reference clip.")
if not consent:
raise PermissionError("Fine print: only clone a voice you own or have permission to use.")
return f"[synthetic voice] {text}"
voicebox = Voicebox()
print(voicebox.clone(3, "This is a short consent-based voice clone demo.", consent=True))
After Cloning: Dictation Hotkey, Whisper Archiving, and Voicebox over HTTP
Once a personality exists, Voicebox's dictation layer turns it into a system-wide input device. Hold a rebindable keyboard chord in any application, speak, and release — the transcript auto-pastes into the field that had focus, with a floating pill showing the recording, transcribing, and refining states . An optional local Qwen LLM cleanup pass strips filler words, fixes punctuation, and removes the repeated loops Whisper sometimes emits; for clips under roughly 5 seconds, supplying a language hint improves accuracy .
Every capture is archived with both audio and transcript, so you can retranscribe the same clip on a different Whisper tier — Base through Large or Turbo — without re-recording .
For agent workflows, Voicebox runs a local MCP server at http://127.0.0.1:17493/mcp over HTTP and stdio, exposing four tools — voicebox.speak, voicebox.transcribe, voicebox.list_captures, and voicebox.list_profiles . Add that URL to the MCP server settings in Claude Code, Cursor, Windsurf, or Cline, and an agent can speak in a cloned personality .
The takeaway: Voicebox is less a single feature than a local pipeline — clone, dictate, archive, and expose over HTTP — with the consent gate staying yours to enforce.
Frequently asked questions
Is jamiepine/voicebox the same product as Meta's Voicebox?
No. Meta's Voicebox was a research model announced in June 2023, trained on more than 50,000 hours of audio and supporting six languages, and Meta declined to release the model or code over misuse risk . The tool covered here is jamiepine/voicebox — an unrelated MIT-licensed desktop app by the author of the Spacedrive file explorer, first released as 0.1.0 on January 27, 2026 . Same name, different lineage.
Does a 3-second audio clip actually work for voice cloning?
Sometimes, but it is below the documented ideal. Some zero-shot engines will process a 3-second seed, yet the official docs recommend 10–30 seconds of clean, low-noise WAV speech in natural, complete sentences . Shorter clips are more sensitive to compression artifacts and background noise, so the "3-second" homepage headline omits the practical trade-off.
Does Voicebox run on Linux?
There is no prebuilt Linux binary as of v0.5.0 (April 25, 2026) — the team attributes this to GitHub runner disk-space limits . Linux users run it via Docker (docker compose up, port 17493) or build from source, which requires Bun, Rust, Python 3.11+, and the Tauri prerequisites .
Can I use Voicebox without an internet connection?
Yes, after the initial model download on first launch. Each TTS engine and Whisper ASR model runs locally once fetched, so generation and dictation work offline . An optional paid cloud backup and sync tier is planned but not active as of v0.5.0 (April 2026) .
How does the cost compare to ElevenLabs?
Voicebox is free for local use under an MIT license . ElevenLabs is a hosted platform: Free at $0 for 10k credits, Starter at $5 for 30k credits (Instant Voice Cloning), and Creator at $22 for 100k credits (Professional Voice Cloning, which needs roughly 30 minutes of audio for fine-tuning) . Voicebox plans a paid cloud backup tier that is not yet live, and output-quality parity with ElevenLabs remains unverified in controlled tests as of April 2026 .
Top comments (0)