Skip to main content
PromptQuorumBuilt for humans. Structured for AI.
Home/Power Local LLM
Voice, Speech & Multimodal

Voicebox Review: Open-Source Local Voice Studio

·7 min readBy Hans Kuepper · Founder of PromptQuorum · Discovery engine for open-weight & open-source AI

Voicebox (version 0.5.0) is an MIT-licensed desktop app that combines voice cloning, text-to-speech and Whisper dictation, and the sources say it keeps your recordings, voice profiles and transcripts on your own computer. It also runs a local API and an MCP server so AI agents can speak in a cloned voice, and every statement here comes from the maker's website, documentation and GitHub repository.

Key Takeaways

  • Maintenance: the main branch last changed on 7 October 2026, so no stale-project warning applies; the latest tagged release, 0.5.0, dates from 25 April 2026.
  • Cost: the local app is free; an optional paid Cloud backup is announced but not yet released.
  • Engines: the README lists eight text-to-speech engines and Whisper for transcription.
  • Platforms: installers for macOS and Windows; Linux needs a build from source.
  • Operator: Spacedrive Technology Inc.; the code is open under the MIT license.

📍 In One Sentence

Voicebox is an open-source desktop app that clones voices, generates speech with local engines, dictates with Whisper and gives AI agents a voice through an MCP server.

💬 In Plain Terms

Think of it as a recording studio and a voice keyboard in one program, with everything running on your own machine instead of a subscription service.

📌Note: This review is based on voicebox.sh, its privacy policy, the documentation and the GitHub repository, checked on 26 October 2026. PromptQuorum has not installed or timed the app, so voice quality and speed are not assessed.

Read in:
🇺🇸en🇩🇪de🇫🇷fr🇯🇵ja🇨🇳zh🇪🇸es🇧🇷pt🇸🇦ar🇰🇷ko

Voicebox is an open-source desktop app that clones voices, turns text into speech with several local engines, dictates into any app with Whisper and lets AI agents speak through an MCP server. Everything it describes runs on your own computer, and the download is free. This review is based on the website, its privacy policy and pricing page, the documentation and the GitHub repository, checked on 26 October 2026; PromptQuorum has not tested the app hands-on.

Voicebox Review: Open-Source Local Voice Studio

What Is Voicebox?

Voicebox is a local-first voice studio that covers both directions of voice: it speaks text aloud in cloned or preset voices, and it listens, transcribes and pastes what you say. The README calls it a free and open-source alternative to ElevenLabs and WisprFlow in one app, and the website describes it as running entirely on your machine.

The app is built with Tauri (Rust) around a Python FastAPI backend, and the website names Spacedrive Technology Inc. as its operator. Version 0.5.0, called "the Capture release" in its notes, added dictation, voice personalities and the MCP server. The version "v0.2.0" shown on the website is a label inside its demo screenshot, not the release number.

Get It

Voicebox is a download for macOS and Windows; Linux has no pre-built binary and must be built from source, and the README also shows a Docker route.

Option
How to get it
macOS.dmg for Apple silicon and Intel: voicebox.sh/download
Windows.msi or .exe installer (x64): voicebox.sh/download
LinuxBuild from source: voicebox.sh/linux-install
Dockerdocker compose up from the repository (README)
Websitevoicebox.sh
Privacy policyvoicebox.sh/privacy
Source codegithub.com/jamiepine/voicebox, MIT license

Version reviewed: 0.5.0, published on 25 April 2026 per the GitHub Releases page, which lists the macOS and Windows installers; the newest change to the main branch was on 7 October 2026. Plans are covered under Plans.

Requirements

No minimum RAM, disk or operating-system version is stated in the sources; the demanding part is the speech and language models, which are downloaded separately.

Operating system

What the sources say:
Installers for macOS (Apple silicon and Intel) and Windows x64; Linux from source; no OS version minimum is given

Acceleration

What the sources say:
MLX on Apple silicon, CUDA on NVIDIA GPUs, ROCm on AMD under Linux, DirectML on Windows, Intel Arc, and a CPU fallback that works everywhere but slower

Lightest engines

What the sources say:
LuxTTS needs about 1 GB of VRAM and Kokoro is an 82-million-parameter model aimed at fast CPU use; the README gives no figure for the larger engines

Models

What the sources say:
Engines, Whisper sizes and the bundled Qwen3 language model (0.6B, 1.7B or 4B) are downloaded inside the app; the sources give no total download size

Permissions

What the sources say:
macOS Accessibility and Input Monitoring for hotkeys and auto-paste; without Accessibility, dictation still saves transcripts and only the paste step is off

Internet

What the sources say:
Needed to download models; the website says generation and refinement then run on the machine

Acceleration backends are as listed in the README; PromptQuorum did not check which combinations ship in the installers.

Features Confirmed by the Sources

Every item below comes from the website, the documentation or the README; none has been independently tested.

  • Voice cloning and presets. Zero-shot cloning from a short reference sample, plus preset voices through Kokoro and Qwen CustomVoice; the README counts eight engines and 23 languages, with the broadest coverage in Chatterbox Multilingual.
  • Expressive speech and effects. Chatterbox Turbo reads tags such as [laugh] and [sigh], and eight post-processing effects from Spotify's pedalboard library can be applied after generation.
  • Long text and stories. Long scripts are split at sentence boundaries and crossfaded, up to 50,000 characters, and a multi-track timeline arranges several voices into a story.
  • Dictation and captures. A global hotkey transcribes with Whisper, or Whisper Turbo, and pastes into the focused field on macOS; every recording is kept in a Captures tab with its transcript.
  • Agents and API. A local REST API on port 17493 and an MCP server with four tools let Claude Code, Cursor, Windsurf or Cline speak in a voice bound to each client.

Plans

The local app is free with no account, and the pricing page lists paid Cloud and Studio plans that are both marked "coming soon". All figures are US-dollar prices from the pricing page, which says they are not final.

Local

What the pricing page lists:
$0, no account: voice cloning across every engine, dictation and captures, MCP and agent integration, personalities, unlimited local generations

Cloud

What the pricing page lists:
$12 per year at launch: end-to-end encrypted backup and sync, 25 GB storage, up to 5 devices, 30-day version history

Studio

What the pricing page lists:
$48 per year, marked as a placeholder: Cloud plus 250 GB storage, unlimited devices, one-year version history, priority support

Figures as shown on 26 October 2026; no other currency or regional price was checked, and taxes are not included.

Data Flow and Privacy

The privacy policy, last updated 6 July 2026, says the desktop app does not transmit your recordings, voice profiles, transcripts, generations or settings to the operator, and it describes no telemetry or analytics for the app. The website adds that models, audio, transcripts and language-model output never leave the machine, and the local API listens on 127.0.0.1 by default.

The optional Cloud backup is described as end-to-end encrypted on your device before upload, so the servers see only the account email, object counts and sizes, hashes, sync counters and timestamps. The policy also covers a Voice ID feature that stores an encrypted voiceprint and warns that published clips are public; payments run through Stripe or by cryptocurrency. The website mentions a "Local or Remote" option to connect to a backend on another machine, which it does not reconcile with its "entirely on your machine" wording.

📌Note: These are the maker's declarations, not audit results. The code is public under the MIT license, but PromptQuorum has not read it or inspected the app's network traffic, and the policy does not address website analytics.

What We Could Not Verify

  • Voice quality and speed. PromptQuorum did not listen to output, measure real-time factors or test cloning accuracy on any engine.
  • Model licenses. The MIT license covers Voicebox's code; each speech model has its own terms, and a recent commit note describes the newly added OmniVoice engine as having non-commercial weights. We did not check any model license.
  • Engine count. The README lists eight engines including Qwen VoiceDesign, while the website lists seven; we cannot tell which is current.
  • Release cadence. The newest tagged release is from April 2026 although the main branch changed in October, so whether the installers include the latest engines is unclear.
  • Installer signing. Whether the macOS and Windows installers are notarized or code-signed was not checked.

Competitors and Alternatives

XTTS v2

Platforms:
macOS, Windows, Linux
Price / license:
Free / CPML
Key difference:
A single cloning model used as a library, with no app, dictation or agent features
Articles about XTTS v2 (10)

+5 more not shown

Coqui TTS

Platforms:
macOS, Windows, Linux
Price / license:
Free / MPL 2.0
Key difference:
A Python toolkit for running text-to-speech models from code rather than a desktop studio
Articles about Coqui TTS (6)

Also mentioned in:

+3 more not shown

Handy

Platforms:
macOS, Windows, Linux
Price / license:
Free / MIT
Key difference:
Dictation only: no speech generation, voice cloning or agent voice output
Articles about Handy (2)

Also mentioned in:

+5 more not shown

OpenWhispr

Platforms:
macOS, Windows, Linux, iOS
Price / license:
Freemium / MIT
Key difference:
Dictation and meeting notes with local or cloud models, and no text-to-speech
Articles about OpenWhispr (1)

These are same-category apps from the PromptQuorum directory; confirm each app's current price, license and platforms on its own page.

Compare in directory

Frequently Asked Questions

Which engine should I start with?

The README describes Kokoro as a tiny 82-million-parameter model with 50 preset voices and fast CPU use, and Chatterbox Multilingual as the broadest for languages, so a CPU-only machine is the case for Kokoro. We did not compare output quality.

Can an AI agent really talk through it?

The README says any MCP-aware agent can call a speak tool, with a per-client voice binding set in Settings, and a plain HTTP endpoint covers tools that do not support MCP. An on-screen pill shows whenever an agent speaks.

Does dictation work on Windows and Linux?

Dictation and the Captures tab are described for all platforms, but the README lists auto-paste on Windows and Linux as roadmap items, so only macOS pasting into the focused field is confirmed.

What does the bundled language model do?

A local Qwen3 model cleans up filler words in transcripts and rewrites text in a voice profile's personality before it is spoken; the README names three sizes, 0.6B, 1.7B and 4B.

Is Voicebox Cloud available?

No: the website shows Cloud as "coming soon" with a notify link, and the pricing page says its prices and limits are not final. The local app does not need it.

Verdict

Voicebox is an actively developed, MIT-licensed way to clone voices, generate speech and dictate on your own computer, with an API and MCP server that let agents speak. The privacy statements are strong for the free local app, but they are the maker's claims, the last tagged release is from April while the code kept changing into October, Linux needs a source build, and auto-paste is confirmed only on macOS. It suits creators and developers who want one free local voice tool and are willing to download several large models; anyone who needs proven voice quality, commercial-use clarity for each model or a polished Linux install should test it and read each model's license first.

Sources

← Back to Power Local LLM

MCP