Orpheus-TTS

What it is

Orpheus TTS is a SOTA open-source text-to-speech system built on the Llama-3b backbone. It provides human-like speech with zero-shot voice cloning, guided emotion and intonation, and low latency streaming (~200ms, reducible to ~100ms with input streaming).

How it works

The project implements TTS models with a finetuned Prod model and a pretrained model for English, plus a multilingual family in a research release. Inference can run on standard setups and through Baseten for one-click deployment with real-time streaming. The README mentions input formatting for prompts and usage with various voices per language, and emphasizes using standard LLM generation parameters (temperature, top_p) with repetition_penalty adjustments for stability.

Getting started

Usage examples and setup paths in the README include:

• Cloning the repo and installing the package:

  git clone https://github.com/canopyai/Orpheus-TTS.git
  

  cd Orpheus-TTS && pip install orpheus-speech # uses vllm under the hood for fast inference
  

  Note: vllm version adjustments may be needed:

  pip install vllm==0.7.3
  

• Streaming inference example script shows how to instantiate the model and generate speech:

  from orpheus_tts import OrpheusModel
  model = OrpheusModel(model_name ="canopylabs/orpheus-tts-0.1-finetune-prod", max_model_len=2048)
  prompt = '''Man, the way social media has...'''
  syn_tokens = model.generate_speech(prompt=prompt, voice="tara")
  

• Colab and Baseten deployment options are listed as one-click deployments.

Recent releases

• The README references a 4/2025 multilingual family release and a 5/2025 Baseten integration, plus a training guide and a language-coverage extension. Specific release notes are not enumerated in the excerpt.

Traction

• GitHub metrics: 6273 stars, 532 forks, 127 open issues. Language: Python. License: Apache-2.0.

Behind the repo

• Baseten is described as the preferred inference partner for one-click production deployment of Orpheus TTS with real-time streaming.

Caveats

• The README notes a known issue with a slightly buggy vLLM version; a workaround is to install vllm==0.7.3 after installing orpheus-speech.
• There are multiple prompts and formatting notes for finetuning and prompting, including voice options and emotive tags. The project includes instructions for finetuning workflows, dataset formats, and training scripts under finetune/train.py.