# Higgs Audio

> Open text-to-speech and voice cloning across 100+ languages from Boson AI

- Category: Multimodal Generation / Audio, Music & Voice
- Language: Python
- Licence: Boson Higgs Audio v3 Research and Non-Commercial License
- Repository: https://github.com/boson-ai/higgs-audio
- Canonical page: https://ai-tldr.dev/tools/higgs-audio/

## In short

Higgs Audio is an open text-to-speech (TTS) model from Boson AI. The latest release, v3, generates conversational speech across more than 100 languages and supports zero-shot voice cloning, so you can match a target voice from a short sample. It sits in the Audio, Music & Voice category of the Multimodal Generation tools and is open-source, written in Python and licensed Boson Higgs Audio v3 Research and Non-Commercial License.

## Overview

Higgs Audio is an open text-to-speech (TTS) model from Boson AI. The latest release, v3, generates conversational speech across more than 100 languages and supports zero-shot voice cloning, so you can match a target voice from a short sample. It also accepts inline controls for emotion, style, and prosody.

It is aimed at developers building voice features who want either a hosted endpoint or full control over the model. You can call the OpenAI-compatible Boson AI API without any GPU, or download the open weights and run them yourself. Both paths cover preset voices, voice cloning, and streaming.

Within the speech and audio space, Higgs Audio sits alongside other neural TTS systems but leans toward expressive, multi-language output. Note that v3 is a standalone release: you grab the weights or call the API directly rather than cloning this repository.

## What it does

- Conversational text-to-speech across more than 100 languages
- Zero-shot voice cloning from a short reference sample
- Inline control over emotion, style, and prosody
- OpenAI-compatible API endpoint, with no GPU setup required
- Open weights you can self-host, with streaming support
- Recommended serving via SGLang-Omni for self-hosted deployments

## Getting started

The quickest path is the hosted Boson AI API, which needs no GPU. You can also download the open weights and serve them yourself.

1. Call the hosted API

   Get a free, rate-limited key at boson.ai/workspace, set it as an environment variable, then post text to the speech endpoint to receive an MP3.

```bash
export BOSON_API_KEY=bai-xxxx

curl https://api.boson.ai/v1/audio/speech \
  -H "Authorization: Bearer $BOSON_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "higgs-audio-v3-tts", "input": "Hello, this is a test."}' \
  --output out.mp3
```

2. Or self-host the open weights

   Download the v3 weights from Hugging Face and serve them with SGLang-Omni. Serving, voice-cloning, and streaming recipes are in the model card and the SGLang-Omni cookbook.

```bash
export HF_TOKEN=hf_xxxxxxxxxxxxxxxx
hf download bosonai/higgs-audio-v3-tts-4b

sgl-omni serve --model-path bosonai/higgs-audio-v3-tts-4b --port 8000
```

## When to use it

- Adding multilingual spoken output to an app or assistant via a hosted API
- Cloning a specific voice from a short sample for narration or characters
- Self-hosting an open TTS model to keep audio generation in your own infrastructure
- Producing expressive speech with controlled emotion, style, and prosody
