Guide
Offline AI voice-over for Reels
Type a script, pick a voice, and get a WAV, an MP3, captions and word timings. It all runs on your own computer once the models have downloaded.
Updated
You can make an AI voice-over for Reels without an account by running a free, open-source text-to-speech engine on your own computer: you type the script, pick a voice and get an audio file, with no API key and no fees per word. The engine's model downloads once, the first time you use it, and after that the voice-over is made offline.
Motion Cheat Code's voice-over tool (the 13-VOICE-OVER folder) puts three such engines behind one command, with 46 voices in 15 languages, and writes files your video project can use straight away. This guide explains what you need and what you get.
Three engines, one tool
Text-to-speech turns written text into speech. The tool uses three free, open-source engines, installed with pip, the Python package installer:
| Engine | Voices | Speed |
|---|---|---|
| Kokoro | English, plus four quick voices for Spanish, French, Italian and Brazilian Portuguese | Fast on any laptop |
| Chatterbox | The other languages | The best quality, but heavy: about 4 GB of RAM and about a minute per sentence group on a normal CPU, much faster on a Mac with Apple Silicon or a PC with an NVIDIA GPU |
| Piper | Ukrainian | Runs fine in a cloud session too |
The languages are English (US and UK), Dutch (Netherlands and Belgium), German, French, Spanish, Italian, Portuguese (Brazil and Portugal), Polish, Swedish, Danish, Finnish, Russian, Greek, Chinese (Mandarin) and Ukrainian. A sample of every voice is in the pack, so you can choose by ear.
Set it up once
You need Python 3.10 to 3.12 and FFmpeg, the free tool HyperFrames also uses to encode video. Install only the engines you need. Their models download themselves from Hugging Face the first time you use a voice; they are not part of the pack, and after that first download the voice-overs are made on your computer.
The simplest route is to let Claude Code do the setup:
Set up the voice-over tool in 13-VOICE-OVER for me.
I only need the English voices for now.
Make a voice-over with one command
python3 vo.py say --voice en-us-female-1 "Type a prompt. Get a reel." --out renders/vo
--voice picks the voice by name, the text in quotes is the script, and --out says where the files go. Three options shape the read:
- Speed from 0.8 (slower) to 1.3 (faster), set with
--speed. - Pauses between sentence groups, longer or shorter.
- Takes. The tool can make several takes and keep the one that Whisper, a speech-recognition model, understands best.
What you get
| File | What it is | Use it for |
|---|---|---|
| WAV | 48 kHz mono, mastered to -16 LUFS, true peak -1.5 dBTP | The voice track in your video or editor |
| MP3 | The same voice-over as a smaller file | Sharing and quick drafts |
| Timings (JSON) | The start and end time of every word and sentence | Timing animation to the voice |
| SRT | Captions | Platforms that accept uploaded subtitles |
| Karaoke transcript | Word timings in the format the karaoke captions template (t32) reads | Word-by-word captions |
LUFS measures how loud audio sounds over its whole length, and true peak is its real highest level, including peaks between digital samples. For the finished social video, aim for about -14 LUFS for the whole mix, with the true peak below -1 dBTP.
Put it in your video
You can ask Claude Code for the whole chain in plain words:
Make a voice-over of script.txt with a US English female
voice, a little faster than normal. Put it in my composition
starting at 0.5 s, add karaoke captions with t32 from the
transcript, and keep the music well under the voice.
In a HyperFrames composition, the voice plays at volume 1.0 and music under a voice at about 0.18 to 0.3. In CapCut, Premiere or DaVinci Resolve, use the WAV or the MP3 like any audio file and keep the music about 14 dB below the voice. Turning the music down while someone speaks is called ducking. For the caption side, see animated captions for Reels.
Credits and honesty
Some voices are built on data under CC BY 4.0, a license that asks for credit. When you publish a voice-over made with one of them, add its credit line; the tool's README lists them. And it is AI speech: say so where a platform or your audience expects it.
On the phone route
If you run Claude Code in the cloud from the Claude app, the English and Ukrainian voices run fine there. The other languages are slow in the cloud without a GPU, so a computer is the better place for them. The phone guide explains that setup.
One real run
In one logged run, a single prompt asked for a 35-second 16:9 launch video with a voice-over for Chairly, a made-up booking app for barbers. Claude Code wrote the script, voiced it with Kokoro, picked music and sound effects from the kit, built and checked every scene and rendered the MP4. It took about 37 minutes from Enter to MP4 after the one-time setup, with no follow-up prompts. Chairly, its numbers and its clients are made up, and times vary with the prompt, the computer and the network.
The voice-over tool is part of the full Motion Cheat Code pack, not the free starter kit. The free kit is still a good way to see how a prompt becomes a video before you add a voice.
Questions
Is the voice-over really offline and free to use?
The engines are free and open source and run on your own computer, with no account, no API key and no fees per word. Their models download once from Hugging Face the first time you use a voice, and after that the voice-overs are made locally. The tool itself comes with the Motion Cheat Code pack.
Which languages can it speak?
The tool has 46 voices in 15 languages, namely English (US and UK), Dutch (Netherlands and Belgium), German, French, Spanish, Italian, Portuguese (Brazil and Portugal), Polish, Swedish, Danish, Finnish, Russian, Greek, Chinese (Mandarin) and Ukrainian.
Do I need a fast computer?
Not for English, because Kokoro is fast on any laptop. Chatterbox, the engine for the other languages, is heavier. It needs about 4 GB of RAM and takes about a minute per sentence group on a normal CPU, and it is much faster on a Mac with Apple Silicon or a PC with an NVIDIA GPU.
Do I have to say the voice is AI?
It is AI speech, so say so where a platform or your audience expects it. Some voices are built on data under CC BY 4.0, and when you publish a voice-over made with one of them, you add its credit line from the tool's README.
Can I make voice-overs from my phone?
On the phone route, Claude Code runs in the cloud. The English and Ukrainian voices run fine there. The other languages are slow in the cloud without a GPU, so a computer is the better place for them.