Orpheus-TTS
Llama-based expressive TTS with emotion tags and ~200ms streaming — human-like speech with guided intonation.
About
Orpheus-TTS is Canopy Labs' 3B-parameter Llama-based speech model (March 2025): text-to-speech with zero-shot voice cloning, guided emotion and intonation via plain-text tags (e.g. <laugh>, <sigh>), and ~200ms streaming latency for realtime use (reducible toward ~100ms with input streaming, per the authors).
Apache 2.0 licensed, with pretrained and finetuned checkpoints on Hugging Face plus multilingual research releases. Inference via the orpheus-speech package (vLLM under the hood). The emotion-tag control is its standout feature versus plainer TTS models. As with all voice-cloning models, impersonation misuse is a real risk — the project disclaims deceptive use.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
No standardized benchmarks. Provider-claimed (Canopy Labs): "superior to SOTA closed-source models" on naturalness — treat as provider-claimed. Independent: broad community adoption (GGUF/ONNX conversions, Baseten one-click deploy) signals real-world usage.
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (metrics tier). Apache 2.0, genuinely useful emotion-tag control, and healthy community adoption (32k HF downloads plus GGUF/ONNX conversions). Docked three points: (1) headline quality claims are provider marketing without independent benchmarks; (2) a 3B Llama-based model needs real GPU resources, unlike lighter TTS options; (3) the project is young (Mar 2025) with a small team behind it.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 6.8/10
- Maintenance · recency, releases, issue load
- 5.0/10
- Openness · license + docs
- 9.1/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · HF downloads 32,128 (peer-normalized)
- · released 574 days ago
- · manual review: emotion-tag control is genuinely distinctive; 3B model needs real GPU; young project, small team; license: Apache-2.0