How to Clone a Voice and Build a Story with Voicebox
Voicebox is a free, open-source AI voice studio that runs entirely on your machine. No cloud, no accounts, no ElevenLabs bill. In this tutorial you’ll install it, clone a voice from 10 seconds of audio, and render a two-speaker story — all local.
TLDR:
- Install Voicebox from GitHub releases (macOS/Windows, ~5 min)
- Clone a voice from 10-30 seconds of clean audio (one drag-and-drop, ~2 min)
- Generate speech from any text in that voice (one click, seconds)
- Build a multi-speaker story in the Stories Editor timeline (drag, position, render)
Prerequisites: macOS 11+ or Windows 10+ with 8GB RAM and 5GB free disk. GPU optional but recommended. Download the right build from github.com/jamiepine/voicebox/releases.
Step 1: Install Voicebox
macOS (Apple Silicon):
# Download and extract
curl -L https://github.com/jamiepine/voicebox/releases/latest/download/voicebox_aarch64.app.tar.gz -o voicebox.tar.gz
tar -xzf voicebox.tar.gz
mv Voicebox.app /Applications/macOS (Intel) or Windows: Grab the matching .msi or .tar.gz from the releases page and install normally.
Launch Voicebox. The first run downloads your chosen TTS engine model automatically (Qwen 1.7B is ~3.5 GB; Kokoro is ~350 MB if you want to start small). Wait for the green status indicator in the bottom-left.
Step 2: Clone a voice from audio
Go to Profiles → + New Profile and pick a cloning engine. Qwen3-TTS 1.7B gives the best overall quality across 10 languages. Chatterbox Multilingual covers 23 languages if you need broader support.
Drag in a WAV file (MP3/M4A/FLAC work too) with 10-30 seconds of clear, quiet speech. No music, no background noise. Click Create.
Voicebox extracts a voice embedding and stores it with the profile. Test it immediately with a phrase like "Hello, this is a test of my voice profile." If it sounds robotic, add a second sample — different tone or speaking style helps.
Pro tip: Record in a quiet room with the mic 6-12 inches from your mouth. Soft furnishings reduce echo. You don’t need a studio — just no fans, AC, or keyboard clatter.
Step 3: Generate speech
Select your new profile from the dropdown in the Generate tab. Type or paste your text. Use normal punctuation for natural pauses. ALL CAPS adds emphasis; italics for subtle emphasis.
Click Generate. The audio appears in seconds (faster with a GPU). Preview, then download as WAV.
For long scripts, Voicebox auto-chunks with crossfade — no manual splitting needed.
Step 4: Build a multi-speaker story
Go to Stories → + New Story. Create one track per speaker.
- Add clips — drag from Generation History, generate new clips inline, or upload audio files
- Position — drag clips onto the timeline, trim edges, adjust spacing
- Render — click the render button to export the mixed audio
The Stories Editor is a lightweight DAW timeline. It’s not full-featured yet (crossfades and effects are coming), but it handles the core use case: multi-host podcasts, audiobook narration with character voices, game dialogue scenes.
Pro tip: Clone two voices — one for each speaker — then generate all clips first, drag them into the timeline, and arrange. You’ll see the whole conversation laid out visually.
Cleanup
Voicebox stores profiles and audio in:
- macOS:
~/Library/Application Support/sh.voicebox.app/ - Windows:
%APPDATA%/sh.voicebox.app/
To reset: delete the profiles folder and restart Voicebox.
If it breaks:
- Model download hangs: Check disk space. Models range from 350 MB to 8 GB.
- Robotic voice: Add more samples. One 10-second clip is the minimum; two clips with different tones is better.
- No audio on macOS: Make sure Voicebox has microphone access in System Settings → Privacy & Security → Microphone.
Related TMFNK Content
- How to Transcribe or Translate Audio Files with Buzz Buzz is the other half of the local voice loop — turn speech into text when Voicebox turns text into speech.
- Set Up LEANN for Private Local RAG on macOS Another local-first AI tool that runs entirely on your machine, no cloud.
- Explore the Vowel Space: Interactive Sound Synthesis A different angle on voice — how vowel sounds actually work and why they matter for synthetic speech.
Crepi il lupo! 🐺