All Things AI · Season 3 · The New Models · Episode 10
OmniVoice is a free, open text-to-speech model that clones a voice from a short sample and speaks any script in it, running entirely on your own computer with no account and no per-use charge. We use it to run our own voices — Al’s and Victoria’s — for every voice clip we send out.
What it is
OmniVoice is a massively multilingual, zero-shot text-to-speech model supporting over 600 languages, built to clone a voice from a short reference sample and produce state-of-the-art quality from that sample alone. It can also design a voice from scratch by adjusting attributes like gender, age, pitch, dialect, or a whisper, without any sample at all, and it supports fine-grained controls like inserting a laugh or correcting how a word is pronounced.
It’s fast enough to run comfortably on ordinary hardware — as low as one-fortieth of the time it takes to speak the audio out loud — and it runs on NVIDIA, Apple Silicon, and Intel graphics hardware, which covers most machines an agent already owns.
How you get to it
You can try it free in a web demo with no install, or run it locally through a Python install or a local web interface once you’ve set it up on your own machine. There’s also a desktop app built specifically to package this kind of local voice work into one interface, with no account and no cloud connection required to use it.
The desktop app
The desktop app wraps the model in a single interface built around real-time dictation, voice cloning, and video dubbing, all running on the desktop with no accounts, no API keys, and no cloud connection at any point. It’s a community project built on top of the underlying model rather than the model’s original maker, which is worth knowing if you go looking for support or documentation.
The controls that make a clone sound natural
Beyond simply cloning a voice, the model supports fine controls that matter for a script that needs to sound like a real person talking rather than a flat read.
Non-verbal sounds can be inserted directly, such as a laugh, so a script doesn’t have to sound uniformly even in tone. Pronunciation can be corrected directly for a name or a term the model would otherwise say wrong, using phonetic spelling rather than hoping the model guesses right. Voice attributes — gender, age, pitch, dialect or accent, even a whisper — can be dialed in individually when you’re designing a voice rather than cloning one.
How we use it
OmniVoice speaks in our own voices — Al’s and Victoria’s, cloned on the same Mac — for every voice clip we send out. Al’s voice was cloned from a set of his own recorded samples, runs locally on the Mac’s own graphics hardware, and speaks a script verbatim and correctly once cloned. Once a voice track exists, it becomes the audio for a talking-head video: a still image made to speak in sync with that same track.
What to use it for
The core use is your own voice, running for free, without sending a script to a paid service every time you need it spoken.
- Cloning your own voice once from a short sample, then reusing that voice for every script afterward at no additional cost.
- Designing a distinct voice from a written description when a clip calls for a voice that isn’t your own.
- Reading a script aloud for a video’s audio track before pairing it with a talking-head image.
- Any routine voice work you’d otherwise pay a subscription for, run instead on a machine you already own.
What it costs
Nothing to download or run. It works entirely on your own hardware, with no account, no API key, and no per-use charge — the only cost is the machine you already have and the time it takes to set it up the first time.
How fast it actually runs
The model’s own documentation reports a real-time factor as low as 0.025 — roughly forty times faster than the length of the audio it’s producing. In practice that means a script that takes a minute to read aloud can come back in a couple of seconds once the model and the voice are loaded, which is fast enough that generating a voice track stops feeling like a step you have to wait on.
Where this goes wrong
One habit worth watching.
Cloning a voice you don’t have clear standing to use. Clone the voices you actually own — your own, a colleague’s with their say-so — rather than a voice you found and liked. For anyone else appearing in a script, a designed voice with no real-person sample behind it is the safer default.
Skipping the design option when a clone isn’t the right fit. If a script calls for a voice that isn’t yours or a colleague’s, describing a voice from scratch avoids the consent question entirely while still giving you a distinct, controllable voice for that clip.
What running your own voice looks like day to day
Once a voice is cloned, the process on any given script is the same every time: run the text through the model on your own machine, get back an audio file in your own voice, and pair that file with whatever video or clip needs a voice track. There’s no per-script charge, no waiting on a queue, and no account to manage — the whole loop lives on the computer that did the cloning.
That reliability is what makes it worth setting up once rather than reaching for a paid service each time. A script speaks correctly and verbatim once the clone is trained, which is the whole point of cloning your own voice in the first place: consistency, at zero marginal cost, script after script.
What to try in the first week
One exercise, built around a voice you already have.
Record a short sample of your own voice, run it through the cloning step once, and then have it read a script you’ve already written for something else. Listen for how close the match is at normal speed before deciding whether this becomes the voice behind your regular video work.
How it sits next to the others
Against a paid voice-cloning subscription, OmniVoice’s whole advantage is that the marginal cost of the next script is zero once the initial clone is made — no per-minute or per-character charge stacking up across a season of video work.
Against a generic preset voice from a video tool, a cloned voice you actually own gives every clip the same recognizable voice across every platform it appears on, which a rotating preset voice can’t do.
Questions agents actually type
Is OmniVoice free?
Yes. It runs entirely on your own computer with no account, no API key, and no per-use charge. The only setup cost is the time it takes to install it the first time.
Can OmniVoice clone a real voice?
Yes, from a short reference audio sample, producing what the model’s own documentation calls state-of-the-art cloning quality. It can also design a voice from scratch, adjusting attributes like gender, age, pitch, and dialect, with no sample needed at all.
Does OmniVoice run on a normal computer?
Yes. It’s built to run on NVIDIA, Apple Silicon, and Intel graphics hardware, and it runs fast enough — as low as one-fortieth of real time — to be usable on ordinary consumer hardware rather than a server.
How is OmniVoice used for real estate video?
The cloned voice becomes the audio track for a script, and that track is then paired with a still image made to speak in sync with it, producing a finished talking-head video without a live recording session.
Can OmniVoice fix how it pronounces a name?
Yes. Pronunciation can be corrected directly using phonetic spelling, so a name or term the model would otherwise say wrong can be locked in ahead of time rather than left to chance.
What languages does OmniVoice support?
Over 600 languages, according to the model’s own documentation, which is what makes zero-shot cloning from a short sample work across a very wide range of scripts and accents rather than a handful of major languages.
Book a 30-minute strategy call →
Season 1 is Claude Code for Real Estate Agents, filmed and on its own hub. All three seasons: AI for real estate agents.