Open-weight model
Orpheus TTS Online
Hear Canopy Labs’ Orpheus without touching a GPU — the Llama-based 3B speech model, running in the cloud. Type text, choose Tara or Dan, and download the audio.
Speech studio
Natural voices, ready in seconds
Standard+ (Orpheus) speaks at a fixed speed — the speed control applies to the other voices.
5,000 free characters/day · No sign-up required · Commercial use included
A language model that speaks
Orpheus takes an unusual route to speech: instead of a purpose-built acoustic architecture, it starts from Llama-3.2-3B and teaches it to emit audio tokens. Delivery inherits the fluency of the underlying language model — pacing, hesitation, emphasis — which is why Orpheus clips tend to sound spoken rather than narrated.
Canopy Labs published the English fine-tune with eight named voices and an Apache-2.0 license, and the model quickly became one of the most-watched open TTS releases. As with its peers, actually running it means a GPU, a serving stack, and patience — which is the step this page removes.
The real model, two of its voices
The generator above serves the actual Orpheus checkpoint with its published Tara and Dan voices — the same names you will find in the model card. If you have seen Orpheus demos and wondered whether they were representative, the honest test is running your own text through it here.
Because it shares a free tier with Kokoro, Chatterbox, and Zonos, you can rerun the same sentence across four open models in a couple of minutes. That side-by-side beats any spec sheet for deciding which one fits your project.
When Orpheus is the right pick
From running the open models side by side:
- Pick Orpheus for conversational lines — dialogue, assistant replies, anything that should sound spoken, not read
- Pick Kokoro for long narration: faster, steadier, and 5,000 characters per clip against Orpheus’s 300
- Self-host Orpheus if you need its emotive tags or the six voices not deployed here