FervorCreative AI
Live Latest 18.11.26 · morning 401 tools tracked 992 workflows indexed 626 topics Hot: Gaussian splatting, ComfyUI, ZONOS3-Mini

Tiny-voca trades the best possible voice for one you can own and ship, an 18 MB English narrator that runs offline in a browser or a game with a clean commercial license, and for a large class of creative projects that trade is the right one.

tiny-vocaQwen3-TTSONNXaudio-genlocal-creative-aiopen-weightslicensing-provenance

Tiny-voca: A Free 18 MB AI Voice You Can Ship Inside Your Game, App or Web Page

A new English text-to-speech model fits in 18 megabytes. That is smaller than a lot of single photos straight off a phone, and it includes everything needed to turn typed text into sound.

Its maker says the whole project cost $9.56 of rented computer time. That figure covers making the training recordings, the failed attempts and the final run. The model is called tiny-voca, it went up on GitHub and Hugging Face on November 14, 2026, and it does one modest thing: it reads English aloud in one of two fixed voices, on an ordinary processor or right inside a web browser, with no internet connection and no bill.

For two years, the voice question creators asked was which service sounds most human. Tiny-voca asks a different question, and for a lot of projects it is the better one: which voice can you actually own and put inside the thing you are making?

Why a small voice matters more than a good one, sometimes

Hosted voice services sound wonderful, and they come with strings. You pay per character. You need a connection. If the service changes its prices, retires a voice or shuts down, your project changes with it. And if you want a voice inside a game, a museum kiosk, an interactive story or a phone app that works on a plane, you need a voice you can bundle with the files.

That is the gap tiny-voca fills. The released model has 4,537,291 parameters, the adjustable numbers a model learns during training. Popular "small" open voice models are much bigger: Kokoro-82M, named for its size, has about 82 million. The author reports tiny-voca reads about 70 times faster than real time on four processor threads, so at that speed a ten-second line takes a fraction of a second.

It runs three ways. There is the original Python version for people comfortable with a command line. There is an ONNX export, a standard model file format that browsers, game engines and phone apps can load. And there is the browser version, which you can try right now at the tiny-voca demo page. The first visit downloads about 40 MB, then you type, pick a voice, set the speed and press Speak.

The license is the other half of the story. Code and model are Apache 2.0, which allows commercial use, modification and bundling. The training dataset is CC BY 4.0. You can ship this inside a paid product.

Where the voices came from

Most free voice tools raise an uncomfortable question: whose voice is this? Tiny-voca has an unusually clean answer. According to the project's model notes, both voices were designed from written descriptions in another open model, Qwen3-TTS, and every training recording was generated by that model. "No recording of a real speaker is involved."

That is a smart move for a project like this. Nobody's voice was copied, and nobody needs to be asked. The two voices, F2 and M2, are American English, one female and one male.

It also sets a ceiling, and the author says so directly. A voice trained only on another model's output "cannot be more natural than the model that generated it." Tiny-voca is a compressed copy of a bigger synthetic voice. It inherits that voice's character and loses some polish on the way down.

How it fits in 18 MB

This is the part worth understanding, because it explains both the size and the sound.

Most modern voice models learn to paint the sound wave itself, sample by sample. That takes a lot of model. Tiny-voca does something closer to how your own voice works. Your vocal cords buzz at a pitch, your breath adds hiss, and the shape of your mouth filters both into vowels and consonants.

Tiny-voca predicts a recipe in that spirit: how long each speech sound lasts, its pitch, how loud it is, whether it is voiced, and the shape of a filter for each tiny slice of time. Then a very small built-in synthesizer makes a buzz and a hiss and shapes them with that filter. The sound-making part of the model has just 25,862 parameters. Almost all the intelligence goes into deciding how each sound should be spoken, and very little into producing the wave.

The result is fast, tiny and clear. The cost is a slightly synthetic edge, like a very good robot-voice effect. The model card lists "slightly synthetic" among its own limits.

What it is good for

Here is where I land: tiny-voca is not a narrator for your documentary. It is a working voice for projects where "free, offline and safe to ship" beats "moving."

  • Game prototypes and jams. Put placeholder dialogue in every character's mouth on day one, then replace the lines that matter with a real actor later.
  • Interactive fiction and web stories. A read-aloud button that costs nothing per listener and works without a server.
  • Kiosks, installations and exhibits. A gallery piece that speaks visitor-typed text, with no network to fail on opening night.
  • Animatics and storyboards. Scratch narration so timing can be judged before a voice session is booked.
  • Accessibility read-aloud on your own site, where clarity matters more than warmth.

One quirk helps here. The same text is always read the same way. For a game that replays a line, that consistency is a feature. For a director who wants five different takes, it is a wall.

Put this into practice

A first session takes about twenty minutes.

  1. Start in the browser. Open the demo and paste three real lines from your project, not a test sentence. Try both voices and a couple of speeds.
  2. Write a torture line. Include a character name you invented, a date, a price and an abbreviation. The model card warns that foreign names and odd spellings are often read wrong, and numbers, abbreviations and words spelled alike but said differently depend on the part that turns text into speech sounds. Find your problem words now.
  3. Respell what it fumbles. If "Aoife" comes out wrong, try typing it the way it sounds. Keep a small list of respellings for your project.
  4. Install the command-line version if the demo convinced you. The README gives a short install (clone, install, download the model, set one path) and a single command that writes a WAV file, with options for voice, speed, pitch and how much the pitch moves.
  5. Split long text into sentences. The model reads one sentence at a time, with a cap of 32 seconds per call. A paragraph should go in sentence by sentence.
  6. Decide between pre-rendering and live speech. For a game with fixed dialogue, render every line once to audio files and ship those. For typed or changing text, bundle the ONNX file and generate on the player's device.
  7. Credit it. Apache 2.0 asks you to keep the license notice, and the pronunciation files carry their own short notices. A line in your credits costs nothing.

The honest limits

English only, two voices, no cloning. You cannot make it sound like you or like your client. If you need a specific voice, this is the wrong tool.

No emotional range. There is no "angry" or "whisper" control. You get speed and pitch adjustments, and the demo page offers only speed.

Its quality scores are tiny. The accuracy check uses 8 to 16 clips, and the author calls it "a sanity check, not a benchmark." No formal listening test has been run.

It will mispronounce things. Invented names, foreign words and unusual spellings are the weak spot.

No streaming yet. It cannot start talking before it has finished a sentence, which matters for live conversational uses.

It is a day old and one person's work. Expect rough edges, and do not build a deadline around it before you have tested your own script.

A voice that belongs to you

The interesting thing about tiny-voca is not how good it sounds. It is how little it takes to own. A voice that fits in a game build, never phones home and was made without borrowing anyone's throat is a different kind of tool from a subscription, even when the subscription sounds better.

Open the demo and paste the line from your project that you have been putting off voicing. If it sounds good enough for a prototype, you just saved yourself a booking. If it does not, you have lost ten minutes and learned exactly what your project needs from a voice.


From the briefing