Back to Gap Map Model components / Speech & audio

Orpheus TTS

Canopy Labs
open weights / Overall score: 1.4

Canopy Labs' text-to-speech model built on a 3B Llama backbone, which generates speech as audio tokens the way a language model generates text. The finetuned English model offers eight voices, takes inline tags for laughs, sighs and other sounds, and streams with about 200 ms latency; the pretrained model, trained on more than 100,000 hours of English speech, clones a voice from example text and audio pairs. A research release adds models for other languages.

Openness

3 medium confidence
3.0
weights
open (finetuned and pretrained 3B checkpoints on the Hub, behind an automatically approved form)
data
described (more than 100k hours of English speech for the pretrained model; not released)
code
open (pretraining and fine-tuning scripts with configs in the repository)
license
Apache-2.0 (declared on both model cards)+Llama-3.2-Community-License (both cards name Llama 3.2 3B Instruct as the base model, and its license travels with derivative weights)
research-release
multilingual-family (2025 research-preview pairs for other languages; the pretrained models are tagged llama3.2 and the finetuned ones apache-2.0)

The weights download after an automatic form and are labeled Apache-2.0, but they are built on Meta's Llama 3.2, whose community license, with its cap on the very largest services, carries over to the model. The training and fine-tuning scripts are published, while the English speech it was trained on is only described.

Adoption

2 high confidence
2.0

Hugging Face downloads of the finetuned and pretrained English checkpoints over the trailing month; the multilingual research models are not counted.

Capability

1 medium confidence
1.0

Orpheus has no listener rating and publishes no results on a public test set, so nothing places it beside the rated synthesizers. It sits with MeloTTS, another popular open voice with no published evaluation.

Verified 2026-11-12