Orpheus 3B

Orpheus 3B is Canopy Labs' open-source text-to-speech system built on a Llama-3b backbone, launched in March 2025. Instead of traditional parametric synthesis, it treats speech generation as a language modeling problem, producing what the team describes as emergent naturalness in intonation, rhythm, and emotion. It was pre-trained on over 100,000 hours of English audio plus billions of text tokens, and is released under Apache 2.0.

Key capabilities include zero-shot voice cloning from a short sample, guided emotion and intonation via simple text tags, and streaming inference with around 200ms latency for realtime dialogue. Canopy later released a multilingual research preview covering additional languages, and partnered with Baseten for optimized fp8 and fp16 inference.