Voicebox

1. What is Voicebox?

Voicebox is a desktop app for working with voices using AI, running on your own computer. It does two main things: it reads text aloud (including cloning a voice from a few seconds of sample audio), and it works the other way round by listening to you speak and typing the words for you. The project's README presents it as an open-source alternative to ElevenLabs (text to speech) and WisprFlow (voice typing).

Voicebox is free and open source (MIT licence). No account is needed. According to the README, the models, voice data and recordings stay on your machine.

A good fit if:

• You make videos, podcasts or audiobooks and want a voiceover without a monthly subscription.
• You want to try cloning your own voice (or the voice of someone who agreed) from a short recording.
• You want to type by voice on Windows or Mac without sending audio to a cloud service.

Not a good fit if:

• You mainly need Vietnamese speech. The README says Voicebox speaks 23 languages but names only some of them, and Vietnamese is not among the named ones. Listen to a test before relying on it.
• Your computer is very weak or your disk is almost full: the app and its models take several GB, sometimes more than ten (see section 2).
• You want to clone someone else's voice without their permission. Use only your own voice or one you have the right to use.

2. What you need

According to the official documentation (https://docs.voicebox.sh):

	Minimum	Recommended
Operating system	Windows 10 or later, macOS 11 or later
RAM	8 GB	16 GB or more
Free disk space	5 GB	10 GB or more
Graphics card	Not required (runs on the CPU)	NVIDIA card with CUDA for faster speech

A few things to know first:

• Running on the CPU works but is much slower. The documentation says roughly 5 to 50 times slower than on a GPU (graphics card).
• The voice models are not inside the installer. The first time you generate speech, the app downloads the model you picked, from about 350 MB (Kokoro) to about 8 GB (TADA 3B). Most people start with Qwen 1.7B (about 3.5 GB). On a slow connection, choose Kokoro or LuxTTS first.
• With an NVIDIA card on Windows, the app also downloads a separate "CUDA backend" when you click install in Settings (about 200 to 400 MB plus about 4 GB).
• No account, no API key and no payment.

The latest release when this was written is v0.5.0 (25 April 2026). There is no ready-made Linux installer yet; this guide covers Windows and Mac only.

3. Step-by-step installation

Step 1: Download the installer

The easiest way is to open https://github.com/jamiepine/voicebox/releases/latest and scroll to "Assets". The home page https://voicebox.sh also has download buttons.

• Windows: download Voicebox_0.5.0_x64-setup.exe (or the .msi if you prefer an MSI installer). It is about 540 MB.
• Mac with Apple chip (M1, M2…): download Voicebox_0.5.0_aarch64.dmg.
• Mac with Intel chip: download Voicebox_0.5.0_x64.dmg.

What you should see: the file in your Downloads folder, around half a GB.

Step 2: Install on Windows

1. Open the .exe you downloaded and follow the installer.
2. If Windows shows a SmartScreen warning ("Windows protected your PC"), click More info and then Run anyway. The documentation explains this is expected because the app is not code-signed. Only do this if you downloaded the file from the official page in Step 1.
3. Open Voicebox from the Start menu.

What you should see: the Voicebox window opens, and the server status dot in the bottom-left corner turns green.

Step 3: Install on Mac

Open the .dmg, drag Voicebox into Applications and open it. If macOS says "App is damaged and can't be opened" (because the app is not signed by Apple), the documentation tells you to clear the quarantine flag with this command in Terminal:

xattr -cr /Applications/Voicebox.app

Step 4 (NVIDIA card on Windows only): turn on GPU acceleration

1. In Voicebox, open Settings and then GPU.
2. If an NVIDIA card is detected, you will see an Install CUDA backend button. Click it and wait for the download (several GB).
3. The app restarts by itself.

What you should see: the GPU section shows the name of your card. If you do not have an NVIDIA card, skip this step; Voicebox picks a suitable way to run on its own.

4. First use

Make your first voice

1. Click the Voices tab in the sidebar, then + New Voice.
2. Give the voice a name and choose its main language.
3. Add a voice sample: click Upload Sample to pick an audio file (WAV, MP3 or M4A), or Record Sample to record live. A good sample is about 10 to 30 seconds of clear speech with little background noise.
4. Click Create Profile.
5. Go to the Generate tab, choose your new voice, type a sentence and click Generate.
6. Click Play to listen and Download to save the file.

What you should see: the first run takes a few minutes because the app is downloading a model (progress is under Settings, Models). Later runs are much faster.

If you have no sample yet, you can use the built-in preset voices of the Kokoro and Qwen CustomVoice models; they need no recording.

Voice typing (dictation)

By default on Windows you hold Right Ctrl + Right Shift, speak, then let go. On Mac it is Right Cmd + Right Option. Tapping Space while holding switches to continuous recording until you tap the chord again. You can change the keys under Settings → Captures → Dictation.

The documentation notes that if the window you want to paste into is running as administrator, Windows blocks a normal-privilege Voicebox from pasting into it.

Speech recognition uses OpenAI's Whisper model running on your machine. The README does not give a language list for this, so test Vietnamese yourself to see whether the quality is good enough.

5. Common problems and fixes

Sign	Cause	Fix
Windows SmartScreen blocks the installer	The app is not code-signed	Click More info, then Run anyway (only if downloaded from the official page)
Red status dot at the bottom left, "Failed to connect to server"	The local server is not running, or port 17493 is taken by another app	Fully close and reopen Voicebox; check your firewall; if the port is taken, close the app using it
First generation takes very long, stuck on "Loading model…"	A model of 350 MB to 8 GB is downloading	Wait for it; follow progress in Settings → Models; on a slow connection pick Kokoro or LuxTTS first
"Out of memory" while generating	The graphics card has too little memory (the documentation says 6 GB+ of VRAM)	Close games or heavy apps; or choose CPU mode in Settings → Generation (5 to 10 times slower); split long text into smaller pieces
The voice sounds robotic or mispronounces words	The sample is short or noisy	Use a clean 10 to 30 second sample, add more samples of the same speaker, write sentences with clear punctuation
Log shows "flash-attn is not installed"	An optional speed-up library is not available on Windows	Ignore it; the documentation says it is harmless

If it still fails, see the official troubleshooting page: https://docs.voicebox.sh

6. Uninstalling and updating

Uninstall: on Windows open Settings → Apps → Installed apps, find Voicebox and remove it. On Mac drag Voicebox out of Applications. Your data (voices you created, generated audio) lives in its own folder; delete it by hand to clean up fully:

• Windows: %APPDATA%\sh.voicebox.app\
• macOS: ~/Library/Application Support/sh.voicebox.app/

Downloaded models take the most space. The documentation does not say where they are stored beyond this data folder, so check the size of that folder after uninstalling.

Update: download the installer of the newer version from the Releases page https://github.com/jamiepine/voicebox/releases/latest and install it over the old one. The documentation says that since v0.4, an upgrade usually only re-downloads a small core of the CUDA backend, not the whole 4 GB.

7. Frequently asked questions

Does it cost money? No. The project is free under the MIT licence, and the README states there are no splash screens, banners or interruptions in the app.

Do I need the internet? Yes for the first run, to download the app and the models. Once the models are downloaded, the README describes a local-first studio where your data does not leave your machine.

Can it run without a dedicated graphics card? Yes, on the CPU, but much slower. A light model such as Kokoro (about 350 MB) is described in the README as fast on a CPU.

Does it speak Vietnamese? The README does not name Vietnamese among its text-to-speech languages. Try a few sample sentences first.

Is voice cloning legal? The project gives no legal advice. Only use your own voice or one whose owner agreed, and label AI-made audio clearly when you publish it.