Orpheus TTS

Overview

Orpheus TTS is an open-source text-to-speech system built on the Llama-3B backbone. Orpheus demonstrates the emergent capabilities of using LLMs for speech synthesis. Check out our blog post to learn more.

Abilities

• Human-Like Speech: Natural intonation, emotion, and rhythm that is superior to SOTA closed source models
• Zero-Shot Voice Cloning: Clone voices without prior fine-tuning
• Guided Emotion and Intonation: Control speech and emotion characteristics with simple tags
• Low Latency: ~200ms streaming latency for realtime applications, reducible to ~100ms with input streaming

Inference

import { OrpheusModel } from "orpheus-speech";

// Load the model
const model_id = "onnx-community/orpheus-3b-0.1-ft-ONNX";
const orpheus = await OrpheusModel({
  model_name: model_id,
  dtype: "q4f16",
  device: "webgpu",
});

// Create a new audio stream
const prompt = "Hey there my name is Tara, <chuckle> and I'm a speech generation model that can sound like a person.";
const voice = "tara";
const stream = orpheus.generate_speech({ prompt, voice });

// Iterate over the audio chunks
for await (const chunk of stream) {
  console.log(chunk.audio.length);
}

// Save the final audio to a file
const result = stream.data;
result.save("output.wav");

Prompting

Voice options (in order of perceived conversational realism): "tara", "leah", "jess", "leo", "dan", "mia", "zac", "zoe". You can also add the following emotive tags: <laugh>, <chuckle>, <sigh>, <cough>, <sniffle>, <groan>, <yawn>, <gasp>.

Additionally, you can use regular LLM generation args like temperature, top_p, etc. We recommend setting repetition_penalty>=1.1 for more stable generation. Increasing repetition_penalty and temperature makes the model speak faster.