- Home
- Text-to-Speech Software
- MARS5-TTS
On the recordRead 12 Oct 202616 facts
MARS5-TTS
Plans and prices
MARS5-TTS
Free
English speech model; AGPL-3.0; commercial licensing inquiries by email
Read 12 Oct 2026 · github.com
On the record
- API access
- The README directs users without the required hardware to use the CAMB.AI API and says they can contact [email protected] about extra test credits.github.com · read 12 Oct 2026
- Commercial use
- Yesgithub.com · read 20 Sep 2026
- Deployment
- The project documents loading through torch.hub or Hugging Face and provides a Docker image and Dockerfile build instructions.github.com · read 12 Oct 2026
- Hardware limit
- Local inference requires GPU capacity to store at least 750M plus 450M parameters and run inference with 750M active parameters.github.com · read 12 Oct 2026
- Inference modes
- MARS5 supports shallow cloning without a reference transcript and deep cloning with one, which typically provides higher quality at the cost of slower inference.github.com · read 12 Oct 2026
- Known limitations
- The project lists inference stability, speed, long-reference selection, and long-form generation among areas needing improvement or contribution.github.com · read 12 Oct 2026
- License
- The English model is open-sourced under GNU AGPL 3.0, and the README says to email CAMB.AI about commercial inquiries or licensing the closed-source version.github.com · read 12 Oct 2026
- Local requirements
- The quickstart lists Python 3.10 or later and Torch 2.0 or later, plus Torchaudio, Librosa, Vocos, Encodec, safetensors, and regex.github.com · read 12 Oct 2026
- Prosody controls
- Punctuation and capitalization in the text can guide pauses and emphasis in generated speech.github.com · read 12 Oct 2026
- Purpose
- MARS5 is CAMB.AI’s English text-to-speech speech model.github.com · read 12 Oct 2026
- Reference audio
- The README recommends clean reference audio between 1 and 12 seconds and says around 6 seconds gives optimal results.github.com · read 12 Oct 2026
- Speech generation
- The model can generate speech from text using a short audio reference, including prosodically challenging scenarios such as sports commentary and anime.github.com · read 12 Oct 2026
- Support
- The repository invites questions and feedback through the CAMB.AI forum and Discord, and lists [email protected] for API credit inquiries.github.com · read 12 Oct 2026
- Technical design
- MARS5 uses a two-stage autoregressive and non-autoregressive pipeline, with a multinomial diffusion component refining speech codes.github.com · read 12 Oct 2026
- Voice cloning
- Speaker identity is specified with an audio reference, and supplying its transcript enables a slower deep-clone mode intended to improve cloning quality.github.com · read 12 Oct 2026
