Voice Cloning Voiceover Tools

Generate voiceovers in your own voice, locally, using Chatterbox TTS by Resemble AI. No subscriptions, no cloud, no usage limits. Everything runs on your machine.

---

How it works

Chatterbox is an open source neural TTS model. You give it a reference clip of your voice and a script, and it synthesizes speech in your voice. It's not an LLM — it's a diffusion-based audio model that captures your vocal identity (accent, cadence, intonation) and applies it to new text.

---

Requirements

1. Install uv

uv is a fast Python package manager. It handles dependencies automatically when you run the scripts — no separate pip install step needed.

Windows (PowerShell):

powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

macOS:

curl -LsSf https://astral.sh/uv/install.sh | sh

Or via Homebrew:

brew install uv

2. Install ffmpeg

Required only if your reference audio is in M4A or MP3 format. Skip if you already have a WAV file or are planning to use WAV only.

Windows:

winget install ffmpeg

Or via Chocolatey:

choco install ffmpeg

macOS:

brew install ffmpeg

3. Make the scripts executable (macOS only)

chmod +x voiceover.py
chmod +x voiceover_server.py

Windows does not use file permissions this way — skip this step.

---

First run

The first time you run either script, Chatterbox's model weights (~3GB total) will be downloaded automatically from HuggingFace. This is a one-time download — subsequent runs load from disk instantly.

• Windows: %USERPROFILE%\.cache\huggingface\hub\
• macOS: ~/.cache/huggingface/hub/

---

Your reference audio

• WAV format preferred. M4A and MP3 also work (ffmpeg handles conversion automatically).
• Aim for at least 30–60 seconds of clean, natural speech.
• Record in a quiet space — no background noise, no music.
• Reading out Harvard sentences makes an excellent reference clip.
• Record your voice and save it as sample-voice.wav in this same directory.

---

Scripts

voiceover.py — single generation

Good for one-off generations. Loads the model fresh each run.

Windows:

# Basic usage
uv run voiceover.py --ref sample-voice.wav --text "Welcome to this video"

# From a text file
uv run voiceover.py --ref sample-voice.wav --file script.txt

# Custom output filename
uv run voiceover.py --ref sample-voice.wav --text "Your script" --out intro.wav

# Tweak voice parameters
uv run voiceover.py --ref sample-voice.wav --text "Your script" --exaggeration 0.4 --cfg 0.7

macOS:

# Basic usage
./voiceover.py --ref sample-voice.wav --text "Welcome to this video"

# From a text file
./voiceover.py --ref sample-voice.wav --file script.txt

# Custom output filename
./voiceover.py --ref sample-voice.wav --text "Your script" --out intro.wav

# Tweak voice parameters
./voiceover.py --ref sample-voice.wav --text "Your script" --exaggeration 0.4 --cfg 0.7

---

voiceover_server.py — persistent server (recommended for sessions)

Loads the model once and keeps it warm. Use this when generating multiple voiceovers in one sitting — each generation is much faster since the model stays in memory.

Terminal 1 — start the server:

Windows:

uv run voiceover_server.py --ref sample-voice.wav

macOS:

./voiceover_server.py --ref sample-voice.wav

Terminal 2 — generate as many times as you want:

# Basic
curl -X POST http://localhost:8765 -d "Welcome to this video"

# Custom output filename
curl -X POST "http://localhost:8765?out=intro.wav" -d "Your full script here..."

# Tweak parameters per request
curl -X POST "http://localhost:8765?exaggeration=0.4&cfg=0.7" -d "Your script"

curl is available natively on both macOS and Windows 10/11.

Output files are saved in your current directory, auto-named voiceover_001.wav, voiceover_002.wav etc. unless you specify ?out=filename.wav.

Long scripts are automatically split into sentence-sized chunks, generated separately, and stitched into one seamless output file.

Stop the server: Ctrl+C in Terminal 1.

---

Parameters

Parameter	Default	Description
--exaggeration	0.5	Expressiveness. Lower = calmer, higher = more animated. Try 0.3–0.7.
--cfg	0.5	How closely the output follows your reference voice. Higher = more like you. Try 0.6–0.7 if it sounds too generic.

Both parameters can also be passed per-request to the server via query string: ?exaggeration=0.4&cfg=0.7

---

Convert output to MP3 or MP4

These commands work on both macOS and Windows (requires ffmpeg installed):

# WAV → MP3
ffmpeg -i voiceover.wav voiceover.mp3

# WAV → MP4 (black screen video, useful for LinkedIn etc.)
ffmpeg -i voiceover.wav -f lavfi -i color=c=black:s=1280x720:r=24 -shortest -c:v libx264 -c:a aac voiceover.mp4

---

Notes

• Device (Windows): If you have an NVIDIA GPU, PyTorch may use CUDA automatically for faster generation. CPU fallback works fine otherwise.
• Device (macOS): Chatterbox runs on CPU on Apple Silicon. MPS (Apple's GPU backend) is not used due to a PyTorch conv1d limitation at this model's output size. CPU on an M-series Mac is fast enough for this use case.
• Watermark: Generated audio includes an imperceptible neural watermark from Resemble AI (Perth watermarker). It does not affect audio quality and is not detected or flagged by YouTube or other platforms.
• Model cache: Safe to delete to free disk space — it will re-download on next run. Windows: %USERPROFILE%\.cache\huggingface\hub\. macOS: ~/.cache/huggingface/hub/.
• Python version: Pinned to 3.11 in the script header. uv handles this automatically.