Orpheus TTS

• 🧠 Llama-3B backbone adapted into an open speech-LLM.
• 🆕 Predicts discrete audio tokens decoded by a SNAC tokenizer.
• 🎯 Supports zero-shot voice cloning and emotion-tag control.
• 🔧 Apache 2.0 license; multiple parameter sizes published.
• ⚡ Streaming-capable design at 24kHz output.
• 💬 Inline emotion tags such as laugh, sigh, yawn, gasp.
• 🏢 Built by Canopy Labs, released on Hugging Face.
• 📚 Pretrained on large volumes of English speech.

canopylabs

Canopy Labs is an AI research group focused on speech and audio generation, working on models that bring natural, expressive voice synthesis to open ecosystems. The team's work centers on text-to-speech systems designed to balance quality, latency, and…

Orpheus TTS is an open-source, Apache 2.0-licensed text-to-speech system from Canopy Labs that reframes speech synthesis as a language-modeling problem. Rather than directly generating waveforms, it is built on a Llama-style 3B backbone and predicts a sequence of discrete audio tokens, which a SNAC tokenizer decodes into 24kHz mono audio. Because it shares the Llama architecture, the model can be served on standard LLM inference stacks.

According to the model card, Orpheus is a Llama-based Speech-LLM designed for high-quality, empathetic text-to-speech generation. The system supports zero-shot voice cloning and granular emotion control through inline tags for cues such as laughter, sighs, coughs, yawns and gasps, allowing developers to steer the expressiveness of synthesized speech.

This entry corresponds to the flagship 3B variant, the finetuned 0.1 release intended for production use. Canopy Labs frames the release as part of a family of differently sized checkpoints, letting developers trade output quality against compute requirements and edge-deployment constraints. The permissive Apache 2.0 license allows commercial use and community finetuning, and the project has accumulated substantial download and like counts on Hugging Face, reflecting broad availability of the open weights.