Orpheus TTS

Orpheus TTS generates speech for developers building voice applications or creating custom voices. Its Llama-based models support local inference and let you control how spoken text sounds.

• Expressive speech: Choose English voices such as tara, leah and leo, and use tags to guide emotion and intonation.
• Voice cloning: The pretrained model can use reference text and speech pairs to reproduce a voice without fine-tuning. It wasn't explicitly trained for zero-shot cloning, and additional reference pairs improve reliability.
• Streaming audio: Generate audio as it becomes available for interactive applications. The streaming package has a known issue that can occasionally skip frames.
• Custom voice training: Adapt models to your own text and recordings with the supplied training scripts, data preparation tools and sample datasets. Fine-tuning supports Hugging Face tooling and compatible approaches such as LoRA.

The project provides pretrained and fine-tuned English models. Local inference uses vLLM, while a llama.cpp implementation offers a CPU-only route. Baseten provides a separate cloud deployment with real-time streaming; Colab notebooks also run on hosted infrastructure.

The code is open source under Apache 2.0. An optional Silent Cipher integration can watermark generated audio.