Robin: A Free AI Voice Model That Trains on Your Own Recordings Instead of Guessing From Ten Seconds
The first line of Robin's documentation is an admission most voice companies would never print: "Zero-shot cloning is never good enough for studio work."
Zero-shot cloning is the trick every voice app sells. Upload ten seconds of someone talking, type a sentence, and hear it back in something like their voice. It demos beautifully. Then you try to use it for a forty-minute audiobook, a game with three hundred lines, or a documentary narration that has to sound like the same person on day one and day nine, and the cracks show. The voice drifts. Certain words come out in somebody else's accent. The laugh sounds borrowed.
Robin, a new voice model you can download and run for free, from Fixed Seed Research Laboratory, a California company, starts from the opposite assumption. It asks you to record fifteen to twenty minutes of clean speech, train a small file on it, and then use that file forever. It runs on your own graphics card, inside the free node-based tool ComfyUI or a simple browser app, and it renders at 48 kHz, the sample rate film and video work in.
That is slower than the apps. My position is that it is also the more honest deal for anyone whose voice work has to survive a client, and that the real thing to scrutinize is not the sound but the license.
What Robin does, in plain terms
Robin is one model with three jobs.
It speaks in a voice you describe. You type a line and a description like "a warm, low voice, close to the mic," and it reads the line in a voice that fits. You can drop cues like [laughs], [sigh], [crying] or [shouting] into the text where they should happen. It handles English, Japanese, or both in one sentence.
It re-performs your own take in a different voice. Record yourself reading a line the way you want it delivered, up to 45 seconds. Robin keeps your words, your timing and your delivery, and swaps the voice. For a director, this is the interesting one. You act the line; the model changes who is speaking. The catch is that you must give it the exact text of what you said.
It learns one exact voice from your recordings. This is the headline feature. You record a set of clips, Robin trains a small add-on file for that one voice, and from then on that file produces that voice on demand. You can dial its strength up or down when you use it. The project ships one such file already: Robin's own character voice.
The model files went up on Hugging Face, the main site where AI models are shared, on November 9 and sat there almost unnoticed; when I checked this morning the model page showed three likes and one download. The code, the ComfyUI nodes and two ready-made workflows went public on GitHub on November 15. This is very new, very small, and not yet tested by many people.
Why training beats guessing
Here is the mechanism, without the jargon.
A ten-second clone works by having the model listen to a short sample and guess what the rest of that person's voice sounds like. It has never heard them ask a question, whisper, get excited or read a number. It fills those gaps with averages from everyone else it has heard. That is why the voice slides around.
Training does the opposite. You give the model twenty minutes of one person saying a wide range of things, and it adjusts a small add-on until it reproduces that range. Nothing has to be guessed at runtime because the model has already heard the questions, the exclamations and the quiet lines.
The project is specific about what "enough" means. It asks for at least fifteen minutes of speech and says its own test voices averaged twenty minutes, spread over about two hundred clips. Each clip should be one or two sentences, three to fifteen seconds long. That is a recording session, not a sample upload.
It also has a sensible idea about checking progress. Training saves a snapshot every 200 rounds and reads a sample sentence in the new voice at each save. You listen through them and pick the one that sounds right. You stop when your ear says stop, which is how voice people already work.
The hardware and the cost
Robin is free to download and run. The money question is your graphics card.
The README lists peak memory use of 8.2 GB to speak, 9.1 GB to change a voice and 9.9 GB to train a new one. On paper a card with 12 GB of memory clears all three, but those peaks were measured on one workstation card, and the project's own web page muddies things: it lists 8 GB in one spot and says training "fits on a single sixteen-gigabyte graphics card" in another. Treat 12 GB as the floor for speaking and changing voices, and 16 GB as the safe bet for training. An 8 GB card is borderline at best.
If you do not have the card, the project page at audiorobin.io offers free test generation through a hosted demo, sponsored by Fixed Seed. That is good for hearing the quality. It will not let you train your own voice, which is the part that matters.
Put this into practice
If you have a 12 to 16 GB card and a quiet room, this is a weekend project.
1. Write the script before you record. The project's best tip is to write the script first and record it line by line, because then the script is already your transcript. Mix statements, questions and exclamations, short lines and long ones. A calm read trains a calm voice, so record the delivery you want back.
2. Record like it is a real session. One speaker, one quiet room with no echo, the same microphone at the same distance throughout. No music or effects. Keep your loudest peaks around minus 6 dB on your meter. Save WAV or FLAC at 44.1 or 48 kHz; the project warns that lower rates train a duller top end, and clips below 16 kHz get left out.
3. Write each transcript exactly. Every clip gets a text file with the same name holding exactly the words spoken. Write numbers the way they are said, "fifty percent," not "50%." If you stumble, re-record the line rather than typing the stumble.
4. Let Robin check your folder. Run the prep step. It lists every clip it will drop and why (no transcript, silent, too short or too long) and flags clipping, digits, low sample rates and transcripts that might not match. Fix those before you spend hours training.
python -m robin prep my_voice/
python -m robin train my_voice/ --model . --out my_voice_lora
python -m robin speak "This is my voice." --model . --lora my_voice_lora/lora.safetensors --out mine.flac
If you live in ComfyUI, the same steps exist as nodes: Robin Check Dataset, then Robin Train LoRA (the node that builds your voice add-on), with a live readout and a playable sample at every save. The repo includes a drag-in workflow file for training and one for speaking.
5. Pick your snapshot by ear. Listen to the sample at each save. Later is not always better. Choose the one that sounds like the person, then test it on lines that were nowhere in your script.
6. Try the voice changer on your own performance. Record a line the way you hear it in your head, then run it through a trained voice. This is where Robin starts to feel like a directing tool rather than a text-to-speech box.
Where it falls short
Nobody has measured it in public yet. There are no published listening tests or comparisons. Everything about quality right now is the makers' word and your own ears.
Two languages. English and Japanese only. If your project needs Spanish or Hindi, look elsewhere.
The voice changer needs perfect transcripts. Robin re-performs a recording only if you give it exactly what was said. Ad-libs and mumbles mean extra typing.
It asks for real effort. Fifteen to twenty minutes of clean, well-transcribed audio is a few hours of work once you count retakes and editing. That is the point, but it is also the cost.
Consent is on you. Robin will train on any clean recordings you give it. Train only on your own voice or a performer who has agreed in writing. The license makes you responsible for having rights to whatever you feed it.
The license is generous with a sharp edge. This deserves its own section.
Read the revenue clause before you sell anything
Robin ships under the Fixed Seed Open License, which is the familiar Apache 2.0 license plus a revenue limit. Personal use, research, education and evaluation are unlimited. Commercial use, including selling the audio you make, is allowed until your "Attributable Revenue" passes two million US dollars in any twelve months.
The phrase to slow down on is "Attributable Revenue." It is not the money the voice earns you. It is the full gross revenue of any product, service or content made with Robin, and the license says that revenue "counts in full and is not apportioned among its parts." Voice a single character in a game with Robin, and the whole game's sales count.
Cross the line and you must tell Fixed Seed in writing within thirty days. Commercial use can continue for ninety days after you first cross the line while you negotiate a separate license, unless Fixed Seed extends that in writing, and after that the license grants no commercial rights. The voice add-ons you train are counted as derivative works, so they stay under the same terms. Fixed Seed claims no ownership of the audio you generate.
For a solo audiobook narrator, a podcaster or a small studio, two million in a year is far away, and this license is a good deal. For anyone building something that might take off, it is a clause to show your lawyer now, not later.
The slower way might be the professional one
Most voice tools are designed to impress in the first minute. Robin is designed for the fortieth minute, the take that has to match the one from Tuesday. It asks you to show up with a script, a quiet room and an afternoon, and in exchange you get a voice file that lives on your own drive and does not change when someone else updates a server.
If you have the card, record twenty minutes this week and train one voice. Use your own. Then listen to it read a line you never recorded and decide whether it would pass in your next project. That answer is worth more than any demo reel, including Robin's.
- Sources: Robin on GitHub · Robin license · Weights on Hugging Face · Project page · Hosted demo