Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

Robin

A 3.58B English and Japanese speech model at 48 kHz that designs voices from text, re-voices your takes and trains a voice LoRA locally.

Reviewed by ReviewRamble editors · Updated

Visit Robin Open source

Share this page:XLinkedInFacebookRedditWhatsAppEmail

Our rating

3.6/5

Robin is a rare open speech model built around training your own voice rather than cloning from ten seconds, held back by two languages, no published benchmarks and a licence with a revenue ceiling.

What is Robin?

Robin is a speech model from Fixed Seed Research Laboratory, the trading name of Fixed Seed LLC, a California company, released on Hugging Face in November 2026. It has 3.58 billion parameters in two parts: a 2.17B autoregressive planner writes the speech as tokens, 25 per second, and a 1.41B flow-matching decoder turns that plan into 48 kHz audio. The project page says it was trained from scratch on 10 million audio clips for 3 days on a single RTX 6000 Pro. It speaks English, Japanese or both in one line, and takes inline tags such as [laughs], [sigh] or [shouting].

It does three jobs. It designs a voice from a written description, it performs a recording of up to 45 seconds again in another voice while keeping the words and timing, and it learns a new voice through a LoRA adapter trained on your own transcribed clips. The package includes a command line, a browser app and seven ComfyUI nodes, including dataset checking and training. Peak GPU memory in bf16 is 8.2 GB to speak, 9.1 GB to change a voice and 9.9 GB to train.

The weights are ungated. On 20 November 2026 the Hugging Face page showed 162 downloads and the GitHub code repository about 15 stars.

Is Robin worth it?

Robin's argument is that zero-shot cloning from a short clip is not good enough for studio work, so it makes LoRA training the main route. The README asks for at least 15 minutes of clean speech; the authors' test voices averaged 20 minutes in about 200 clips of 3 to 15 seconds. Training saves a checkpoint every 200 steps and reads a sample line at each one, so you choose the step by ear. The project page says a voice like Robin's own fits on a 16 GB card, and the published peak is 9.9 GB.

The voice changer is the less common feature. Given a take of up to 45 seconds and its exact transcript, it performs the same line in another voice with the original timing and delivery, which suits replacing a scratch read or a placeholder performance. On NVIDIA GPUs, CUDA graphs and merging the LoRA into the weights make speaking with a LoRA about 4 times faster, according to the README.

The licence is where buyers need to read closely. The Fixed Seed Open License 1.0 is Apache 2.0 with additions. Commercial use of the model, anything derived from it and its audio is free while your Attributable Revenue stays at or below US$2,000,000 in any 12 consecutive months. That revenue is gross, counts the whole product that uses Robin's audio rather than a share of it, and includes every company in your group. Once you pass it you must tell Fixed Seed within 30 days, may continue for 90 days while negotiating, and then need a separate written agreement. Personal use, research, education and evaluation are not limited. Trained LoRAs count as derivative works and carry the same terms, and the licence is governed by California law. The CosyVoice 3 speech tokenizer bundled with it stays under Apache 2.0.

The catch: Fixed Seed publishes no word-error rates, speaker-similarity scores or listening tests, only audio samples, so quality claims cannot be checked against VoxCPM2 or Qwen3-TTS. It covers two languages where rivals cover dozens. The README describes no watermark or consent check, and the licence puts responsibility for having rights to the recordings you train on with you. The project is under two weeks old with about 15 stars.

Our take: for a small studio or solo creator working in English or Japanese who can record 20 minutes of a voice they have rights to, Robin is worth an evening's test in ComfyUI. Teams that need many languages or audited numbers are better served by VoxCPM2's Apache 2.0 licence.

Where it earns its place

  • An indie game developer voicing a lead character

    Keep one consistent voice across hundreds of lines without booking repeat sessions

    Record about 20 minutes with the actor's written consent, run robin prep and robin train on a 16 GB card, and render lines with the LoRA. Free while the studio stays under US$2,000,000 in revenue in 12 months.

  • A video editor at a small agency

    Replace a scratch narration with the final voice while keeping the timing

    Feed the scratch take, up to 45 seconds at a time, and its exact transcript to the Robin Voice Changer node with the target LoRA or a described voice. Edit points survive because timing and delivery are kept.

  • A Japanese VTuber producing bilingual clips

    Read English and Japanese in one line in a single character voice

    Train a LoRA on Japanese and English recordings of the character, then speak with --lang mix. Inline tags such as [laughs] add delivery cues without a separate pass.

Before you commit

Setup
hf download FixedSeed/Robin, then pip install the bundled package (add the gui extra for the browser app); or copy comfyui-robin into ComfyUI's custom_nodes, which downloads the weights on first run.
Integrations
Python command line, a local browser app and ComfyUI nodes for loading, LoRA, speaking, voice changing, dataset checks and training. No first-party API; Lattice offers free test inference as a sponsor.
Data and privacy
Self-hosted, so text and recordings stay on your machine. The licence makes you responsible for rights to the text and recordings you supply; no watermark is described.
Cost at scale
No fee below the revenue threshold of US$2,000,000 in any 12 consecutive months; above it, a separate commercial licence with terms not published. Otherwise the cost is your own GPU time.
Limits
English and Japanese only; voice changing up to 45 seconds per take and needs an exact transcript; training clips of 1 to 30 seconds, below 16 kHz left out. No benchmark figures published.

What’s good

  • LoRA training, dataset checks and ComfyUI nodes included from day one
  • 48 kHz output; peak 8.2 GB to speak and 9.9 GB to train
  • Re-voices a take of up to 45 seconds keeping its timing
  • Commercial use free below US$2,000,000 a year in revenue

What’s not

  • No published benchmarks, only samples
  • English and Japanese only
  • Revenue ceiling counts a whole product's gross revenue
  • No watermark or consent check described

Best for

Small studios and independent creators in English or Japanese who want a dedicated voice trained on their own recordings and run on one GPU.

Robin alternatives

Robin: common questions

Is Robin worth it?
Robin is a rare open speech model built around training your own voice rather than cloning from ten seconds, held back by two languages, no published benchmarks and a licence with a revenue ceiling. We rate it 3.6 out of 5.
How much does Robin cost?
Free under the Fixed Seed Open License 1.0, commercial use included until your Attributable Revenue passes US$2,000,000 in any 12 consecutive months; above that a separate licence is needed. No first-party paid API.
What are the downsides of Robin?
No published benchmarks, only samples; english and Japanese only; revenue ceiling counts a whole product's gross revenue; no watermark or consent check described.
Who is Robin best for?
Small studios and independent creators in English or Japanese who want a dedicated voice trained on their own recordings and run on one GPU.
How much does Robin cost for a team or in production?
No fee below the revenue threshold of US$2,000,000 in any 12 consecutive months; above it, a separate commercial licence with terms not published. Otherwise the cost is your own GPU time.
Is Robin safe for company data?
Self-hosted, so text and recordings stay on your machine. The licence makes you responsible for rights to the text and recordings you supply; no watermark is described.
What are the best Robin alternatives?
The closest alternatives we have reviewed are VoxCPM2, Qwen3-TTS, Fun-CosyVoice3 and Higgs TTS 3.

Answers checked Nov 20, 2026.