Zero-shot cloning is never good enough for studio work. That's why we trained Robin: an open-weights model built for professionals in the open-source community. It launches with LoRA training scripts included and it renders at 48 kHz, the standard for film and video. Design a voice through text, re-voice your own takes with another voice or train a dedicated adapter model to always produce an exact voice using your own voice data.

English and Japanese, or both in one line
48 kHz audio
3.58 B parameters: a 2.17 B planner writes the speech as tokens, and a 1.41 B flow-matching decoder turns that plan into sound
Peak GPU memory in bf16: 8.2 GB to speak, 9.1 GB to change a voice, 9.9 GB to train a voice

Files

File
`robin_v1.safetensors`, `robin_vae.safetensors`, `config.json`, `tokenizer.json`	the model
`speech_tokenizer_v3_25hz.onnx`	CosyVoice 3's speech tokenizer, for long text, voice changing and training
`robin_character_lora.safetensors`	Robin's own voice, as a LoRA
`package/`	the `robin` Python package: command line, app and LoRA training
`comfyui-robin/`	the ComfyUI nodes
`comfyui-workflow-examples/`	ComfyUI workflows to start from

Install

hf download FixedSeed/Robin --local-dir robin
cd robin
pip install ./package
pip install "./package[gui]"   # for the app

Use

python -m robin speak "That went better than expected [laughs]" --model . --voice "a warm, low voice, close to the mic" --out take.flac
python -m robin speak "Exactly what my take says." --convert my_take.wav --model . --voice "a bright, fast-talking woman" --out revoiced.flac
python -m robin gui --model .

Languages: --lang en (default), --lang ja, or --lang mix for English and Japanese in one line.

Inline tags in square brackets, such as [laughs], [sigh], [crying] or [shouting].

--convert performs a recording of up to 45 seconds again in another voice, keeping its words, timing and delivery. The text must be exactly what the recording says.

robin_character_lora.safetensors is Robin's own voice:

python -m robin speak "Hi, I'm Robin." --model . --lora robin_character_lora.safetensors --out robin.flac

The app runs in your browser: speak with two takes side by side, change a voice from a description, and train a voice with live progress and a sample at every save.

Train a voice

Robin learns a new voice through a LoRA trained on transcribed recordings of it. Each voice is saved as one file, and you set its strength when you use it.

A dataset is a folder of clips, each with a .txt transcript of the same name, plus an optional voice.txt that describes the voice:

my_voice/
├── line_001.txt    # what line_001.wav says
├── line_001.wav
├── line_002.txt
├── line_002.wav
├── …
└── voice.txt       # optional: the voice in plain words

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

prep checks the folder first. It lists the clips it will leave out (no transcript, silent, too short or too long) and notes anything worth fixing: clipping, digits, low sample rates, transcripts that may not match, repeated lines. train saves a checkpoint every 200 steps and reads a sample line in the new voice at each save, so you can pick a step by ear. Stop it at any time; running it again continues from the last save.

Recommendations:

How much: at least 15 minutes of speech. Our test voices averaged 20 minutes, in about 200 clips.
Each clip: one or two sentences, 3 to 15 seconds. Anything from 1 to 30 seconds is accepted, so split longer takes.
Recording: one speaker in a quiet, dead-sounding room, with the same mic and distance throughout. No music, effects or other voices. Keep peaks around −6 dBFS.
Format: WAV or FLAC at 44.1 or 48 kHz. Lower rates train a duller top end, and clips below 16 kHz are left out.
Transcripts: exactly the words spoken. Write numbers the way they're said ("fifty percent"), and mark sounds and delivery with tags like [laughs] or [sigh] where they happen. Re-record a stumble instead of transcribing it.
Range: mix statements, questions and exclamations, short lines and long ones. A calm read trains a calm voice, so record the delivery you want back.
Tip: write the script first, then record it line by line. The script is the transcript.

ComfyUI

Copy comfyui-robin into ComfyUI/custom_nodes and install its requirements into ComfyUI's Python:

pip install -r ComfyUI/custom_nodes/comfyui-robin/requirements.txt

The first run downloads Robin from this repo: the model into ComfyUI/models/robin/robin, and Robin's own voice into ComfyUI/models/loras. To place the files yourself, put robin_v1.safetensors, robin_vae.safetensors, speech_tokenizer_v3_25hz.onnx, config.json and tokenizer.json in ComfyUI/models/robin/robin, and robin_character_lora.safetensors in ComfyUI/models/loras.

Two workflows to start from. Drag one onto the ComfyUI canvas:

Robin - LoRA Voice Inference Official.json: Robin's own voice reads a line.
Robin - LoRA Voice Training Official.json: checks a folder of clips and trains a voice LoRA on it.

Node
Robin Loader	Loads Robin, downloading whatever is missing on the first run. Its readout shows what is on disk, what is downloading and when Robin is ready.
Robin LoRA	Applies a voice LoRA at any strength from 0 to 2. `robin_character_lora.safetensors` is Robin's own voice; the voices you train are listed under `robin/`.
Robin Speak	Reads text in a voice you describe, or in the LoRA's voice. Long text is read in steady pieces, and tags like `[laughs]` can go anywhere. The last take plays on the node, with a download button; connect Save Audio to keep every take.
Robin Voice Changer	Performs a recording of up to 45 seconds again in another voice, keeping its words, timing and delivery. Give it exactly what the recording says.
Robin Load Audio	The recording for the Voice Changer. Record takes a line from your microphone and trims the silence around it; Choose file, or dropping a file on the node, adds one from your computer. Both are kept in `ComfyUI/input`. Recording works when ComfyUI is open on localhost or over https.
Robin Check Dataset	Checks a folder of clips before training. Choose folder uploads one from your computer into `ComfyUI/input/robin_datasets`, or type a path on the machine running ComfyUI. The report lists the clips it will leave out, and why.
Robin Train LoRA	Trains a voice LoRA on the checked folder, with live steps, losses and time left. Every save writes a checkpoint and reads the sample line in the new voice, which you can play on the node. Ranks, learning rates and the other settings fold away under advanced. Cancel the queue to stop; queue it again to continue from the last save.

Trained voices are saved in ComfyUI/models/loras/robin/<name>: the finished lora.safetensors, and a checkpoint for each save.

Speech tokenizer: CosyVoice 3.

License

Robin is free to use, change and build on, commercially too, under the Fixed Seed Open License, which is based on Apache 2.0. If your revenue from Robin and what you make with it passes US$2 million in a year, write to us at support@fixedseed.com about a commercial license. Questions about Robin go to the same address.

Downloads last month: 1