Orpheus TTS
Open-weight, human-sounding TTS with zero-shot voice cloning and low-latency streaming, built on Llama-3b.
Overview
What is Orpheus TTS
Orpheus TTS is an open-weight, Llama-based text-to-speech model released in March 2025 by Canopy AI. It delivers human-like speech with natural intonation and emotion, supports zero-shot voice cloning, and enables guided prosody control, all with low-latency streaming suitable for real-time applications.
Running it privately on Venice
On Venice, Orpheus TTS runs with zero retention — your text inputs and generated speech are never stored or profiled. This ensures private, sovereign TTS generation, ideal for sensitive or conversational use cases where user privacy is paramount. You retain full control over voice data.
Assessment
Strengths and limitations
- Human-like speech with natural rhythm, emotion, and expressiveness competitive with closed-source leaders.
- Zero-shot voice cloning: generate speech in a target voice without fine-tuning.
- Guided emotion and intonation control via simple input tags.
- Low-latency streaming (~200ms, reducible to ~100ms) ideal for real-time interactive applications.
- Open weights under permissive licensing (code: Apache 2.0, weights: Llama-derived)
- Model weights require agreement to Llama license terms, limiting full open redistribution.
- Voice cloning quality depends on reference audio quality and length.
- No end-to-end encryption or TEE protection on Venice, though prompts are not retained.
Samples
Sample outputs
Generated on Venice with our standard prompt suite — the same scripts we run through every model of this type, so you can judge it like-for-like.
“On Venice, your prompts are processed privately and never stored, profiled, or used to train anyone's model.”
“Wait — so I can run a private voice model with zero data retention, and pay only for what I use? That's genuinely useful.”
“Three… two… one… liftoff! The rocket roared into the night sky as the crowd erupted in cheers.”
Specifications
Datasheet
- Maker
- Canopy AI
- Released
- March 2025
- Architecture
- Llama-3b-based Speech-LLM
- Parameters
- 3B
- Open weights
- Yes
- Privacy on Venice
- Private — zero retention
- Available on Venice since
- Apr 2026
- License
- Apache 2.0 (code), Llama-derived (weights)
API
Call it from your code
Venice exposes an OpenAI-compatible API. Point your base URL at Venice and pass the model id.
curl https://api.venice.ai/api/v1/audio/speech \
-H "Authorization: Bearer $VENICE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "tts-orpheus",
"input": "On Venice, your prompts are processed privately.",
"voice": "af_sky",
"response_format": "mp3"
}' --output speech.mp3Pricing
What it costs on Venice
Billed per character on Venice: $62.50 per 1M characters of synthesized speech.
New Venice accounts include a free daily allowance and 500 welcome credits — no credit card required.
Alternatives
How it compares
| Model | Best for | Price | Voice cloning | Open weights | Latency |
|---|---|---|---|---|---|
| Orpheus TTS | Open-weight leader in expressiveness and cloning flexibility. | $62.50 / 1M chars | Yes — zero-shot | Yes | ~200ms |
| Chatterbox HD (Resemble AI) | Slightly cheaper but closed weights; strong cloning. | $50 / 1M chars | Yes | No | — |
| ElevenLabs Turbo v2.5 | Comparable price and quality, but not open or private by default. | $62.50 / 1M chars | Yes | No | — |
| Kokoro Text to Speech | Far cheaper and open, but lower expressiveness and cloning capability. | $3.50 / 1M chars | Limited | Yes | — |
Open-weight leader in expressiveness and cloning flexibility.
Use cases
What it is good for
- 01Privacy-preserving voice assistants and chatbots with expressive, human-like responses.
- 02Real-time dubbing and voiceovers with controlled emotion and tone.
- 03Custom voice cloning for content creators and game NPCs without retraining.
- 04Multilingual applications using the research-preview multilingual variants.
- 05Applications requiring on-premise or sovereign TTS without cloud leakage.
Prompting
Getting better results
Use emotion tags (e.g., [happy], [serious]) to guide prosody and intonation.
Provide high-quality, clear reference audio for zero-shot voice cloning.
Stream input text to reduce latency in real-time applications.
Fine-tune the base model for domain-specific voices or accents using the provided training code.
Version history
Base model release
Initial finetuned version
Multilingual variants released
FAQ
Frequently asked questions
Orpheus TTS is an open-weight, Llama-based text-to-speech model developed by Canopy AI and released in March 2025. It generates human-like speech with natural emotion, rhythm, and zero-shot voice cloning, optimized for real-time, expressive applications.
On Venice, Orpheus TTS is priced at $62.50 per 1 million characters of synthesized speech. There are no subscription fees — you pay only for what you generate.
Orpheus TTS is open-weight: the code is Apache 2.0 licensed, and the model weights are available under a Llama-derived license. You can use, modify, and deploy it, but redistribution is subject to Llama terms.
Yes. Orpheus TTS supports zero-shot voice cloning — you can generate speech in a target voice using a short reference audio clip without any fine-tuning required.
Yes. You can guide emotion and intonation using simple text tags in the input, allowing expressive control over the generated speech for applications like storytelling or customer service.
Orpheus TTS has a streaming latency of approximately 200ms, which can be reduced to around 100ms with input streaming — making it suitable for real-time conversational interfaces.
Orpheus TTS matches ElevenLabs Turbo in voice quality and cloning ability while offering open weights and privacy on Venice. ElevenLabs is closed and may profile users; Orpheus is better for sovereign, uncensored deployment.
Yes. Orpheus TTS has been available on Venice since April 2026, running with zero retention — your prompts and audio are never stored, ensuring full privacy and compliance.
Run Orpheus TTS privately
No prompt logging. No data used for training.