🍋 Higgs Studio

Local, expressive text-to-speech with a real expression console.

Zero-shot voice cloning · 100+ languages · inline control over emotion, style, prosody & sound effects — all running on your own GPU.

A clean Gradio front-end for Higgs Audio v3 TTS by Boson AI.

---

⚠️ Read the Responsible Use section before you do anything. Higgs Studio can clone a human voice from a few seconds of audio. That power comes with rules — legal ones and decent-person ones. This tool is for your own voice and voices you have explicit permission to use. Nothing else.

---

Why this exists

Most local TTS front-ends give you a single text box and call it a day, which means the model's best trick — controlling how a line is delivered — stays buried. Higgs v3 can shift emotion, style, pitch, speed, pauses, and drop in sound effects mid-sentence, all from inline tags. Higgs Studio surfaces every one of those in the UI so you can actually use them, then runs the whole thing locally with no quotas, no per-character billing, and no audio leaving your machine.

Features

• 🎙️ Expression console — emotion, style, and prosody pickers plus one-click sound-effect and pause buttons that inject the right control tokens for you.
• 🗣️ Zero-shot voice cloning — drop in a reference clip; the app auto-transcribes it on CPU to sharpen fidelity (consent required — see below).
• 🌍 100+ languages — single-digit WER/CER across 85 production-quality languages.
• 🖥️ Fully local — runs on your CUDA GPU. No ZeroGPU quotas, no API keys, no telemetry.
• 🎛️ Sampling control — temperature, top-p, top-k, seed, max-tokens, all exposed.
• 💅 Glassmorphism UI — tabbed layout, system-status panel, streaming model downloader.

Quick start

git clone https://github.com/sjfischr/higgs-studio.git
cd higgs-studio
pip install -r requirements.txt        # install the CUDA build of torch from pytorch.org
python app.py

Open the ⚙️ Setup tab → Download all models (~9.4 GB), wait for it to finish, then head to 🎙️ Studio.

A note on the model repo: Higgs Studio downloads the transformers-ported build multimodalart/higgs-audio-v3-tts-4b-transformers, which is what exposes generate_speech() to plain 🤗 Transformers. Boson's raw bosonai/higgs-tts-3-4b weights are built for vLLM-Omni / SGLang-Omni serving and won't drive this app directly. Same v3 model, different packaging.

Hardware

GPU	NVIDIA with ~12 GB+ VRAM (runs comfortably real-time on an RTX 4090)
Disk	~12 GB free for model weights
OS	Linux / Windows with a CUDA-enabled PyTorch

Expression tokens

The pickers in the UI insert these for you, but you can also type them straight into the text box. All tags use <|category:value|> syntax and can be placed mid-utterance.

Placement rules that matter: put delivery tags (emotion / style / prosody) at the start of a turn; put <|prosody:pause|> / <|prosody:long_pause|> exactly where the pause should land; and pair every <|sfx:…|> with its onomatopoeia immediately after (e.g. <|sfx:laughter|>Haha). Piling many tags head-to-head with no text between them can make generation wander — lead with intent, then let the words carry it.

Emotion (21)

elation · amusement · enthusiasm · determination · pride · contentment · affection · relief · contemplation · confusion · surprise · awe · longing · arousal · anger · fear · disgust · bitterness · sadness · shame · helplessness

Style (3)

singing · shouting · whispering

Sound effects (9) — pair with onomatopoeia

Token	Say right after
sfx:cough	Ahem
sfx:laughter	Haha / Hehe
sfx:crying	Boohoo / Sob
sfx:screaming	Ahh / Aaah
sfx:burping	Burp
sfx:humming	Hmm / Mmm
sfx:sigh	Uh / Ahh
sfx:sniff	Sff
sfx:sneeze	Achoo

Prosody (10)

speed_very_slow (≈0.65×) · speed_slow (≈0.85×) · speed_fast (≈1.2×) · speed_very_fast (≈1.4×) · pitch_low (≈−3 st) · pitch_high (≈+2.5 st) · pause (≈400–700 ms) · long_pause (≈700–1500 ms) · expressive_high · expressive_low

Example:

<|emotion:amusement|><|prosody:expressive_high|>Wait, that was kind of hilarious. <|sfx:laughter|>Hehe, I was not ready for that.

🛡️ Responsible Use & Consent

Higgs Studio synthesizes and clones human voices. Read this. It's short and it's not optional.

Only clone voices you're allowed to clone.

• ✅ Your own voice.
• ✅ A voice you have explicit, documented permission to use.
• ❌ Anyone else. Full stop.

Never use Higgs Studio to:

• Impersonate a real person to deceive, defraud, or take action on their behalf.
• Create voice content of someone without their consent — especially anything sexual, harassing, or defamatory.
• Generate political or election content designed to mislead.
• Produce robocalls, scam calls, or any communication that pretends to be someone it isn't.
• Clone the voice of a minor, for any reason.

The law is real and getting realer. Voice is now a protected right in a growing number of places. Tennessee's ELVIS Act makes unauthorized voice cloning actionable and explicitly reaches the distribution of cloning tools; the federal NO FAKES Act (advancing through Congress as of mid-2026) would create liability for producing or distributing unauthorized digital voice replicas. "I only ran the software" is not the safe harbor people assume it is. Know your local law before you generate.

Disclose AI-generated audio when you share it, so listeners aren't deceived.

This project ships no watermarking or identity verification — it's a local research tool, not a managed platform. That puts the entire burden of responsible use on you. If you can't carry that, don't use it.

Built something cool and consensual with it? That's the point. Built something that hurts someone? That's on you, and the maintainers want no part of it.

License & attribution

• Higgs Studio code: [MIT] © its contributors.
• Higgs Audio v3 model: governed by the Boson Higgs TTS 3 Research and Non-Commercial License. Personal, non-commercial, self-hosted use only. Commercial, hosted, or revenue-generating use requires a separate license from Boson AI. This repo is non-commercial by design.
• Original Gradio app: adapted from multimodalart/higgs-audio-v3-tts.
• Moonshine ASR: UsefulSensors/moonshine-base, used for reference-clip transcription.

By using this software you agree to the model license and to the responsible-use terms above.

Acknowledgments

Higgs Audio v3 by Boson AI · transformers port by multimodalart · ASR by Useful Sensors.

---

_{Higgs Studio is an independent, non-commercial wrapper. Not affiliated with or endorsed by Boson AI.}