Orpheus
A Llama-based open-weight-tts family from Canopy AI (Mar 2025), Apache-2.0 licensed, shipped in 3B / 1B / 400M / 150M sizes open-source-tts-models.
Profile
- Capabilities: zero-shot voice-cloning, guided emotion, and real-time streaming — an expressive, controllable all-rounder.
- Multilingual variants (Chinese, Hindi, Korean, Spanish).
- Size ladder: the 150M–400M tiers target on-device/low-latency; the 1B–3B tiers target quality — one family spanning the efficiency/quality range.
Why it matters here
Orpheus is a prime example of the TTS-on-an-LLM-backbone pattern: built on Llama, it inherits text-LLM architecture for speech (cf. neural-audio-codec; bridge to llm-providers-wiki’s gemini/open-weight-Llama thread). Permissive Apache-2.0 + cloning + streaming makes it a common production pick where fish-audio-s2-pro‘s research license is a blocker.
Related
open-weight-tts · text-to-speech · voice-cloning · sesame-csm · neural-audio-codec · open-source-tts-models