🚀 One-click cross-platform install via Pinokio:

Works on Windows / Linux (x64 & aarch64) / macOS · NVIDIA / AMD / Apple Silicon / CPU. No install.bat — Pinokio sets up CUDA, Python 3.12, PyTorch and dependencies for you. Launcher: timoncool/HiggsAudio-Studio-pinokio

Generate expressive speech in 100+ languages, clone any voice from a reference clip, let a local LLM direct the delivery, and produce full podcasts and audiobooks — entirely on your machine. 100% offline, no cloud, no API keys. Everything lives inside the folder: Python, dependencies, models, cache. Delete the folder — the app is gone.

Built on Higgs Audio v3 TTS by Boson AI — a 4B expressive TTS model with native multi-speaker and inline emotion/prosody control.

Features

• 🎙️ TTS — text → speech, 100+ languages; temperature, top-p/k, seed lock; autoplay; ⏹ Stop interrupts generation on the fly (at the inference level); per-frame progress in the terminal.
• 🎭 Expression + 🤖 Director — insert control tags (<|emotion|>, <|sfx|>, <|prosody|>, <|style|>) with buttons + an Enrich button: a lightweight local LLM normalizes the text and places emotion/sound/prosody tags by meaning.
• 🧬 Voice cloning — zero-shot from a reference clip with auto-transcription (Moonshine ASR); preset library + on-demand Russian voice pack.
• 🎬 Podcast / Dialogue — the LLM writes a multi-speaker script, each speaker its own voice → stitched with loudness leveling across speakers (LUFS −16, podcast standard — no speaker quieter than another).
• 📚 Audiobook — narrator/character attribution with a persistent roster (same character = same voice), long-form with timbre carry-over + loudness normalization.
• 📦 Batch — a list of texts → mass synthesis with a live log.
• 💾 Output format — WAV / MP3 / FLAC / OGG; results saved to output/ with timestamps.

43 control tags: 21 emotions, 10 prosody, 3 styles, 9 sounds. AI director — switchable GGUF model: Qwen3.5-9B (default, ~5.5 GB) / Qwen3.5-4B (light, ~2.5 GB), runs on GPU via llama.cpp. Higgs TTS quantization is on-the-fly (⚗️ experimental), auto by VRAM. RU / EN UI.

System Requirements

Platforms (via the Pinokio launcher)

Higgs uses PyTorch SDPA (flash kernels built in) and does not need external Flash-Attention 2. The local install.bat build targets NVIDIA Windows; full cross-platform support is via Pinokio.

Memory (NVIDIA; TTS is quantized on the fly, the LLM director loads separately)

Models (~9 GB TTS + LLM) download automatically on first run.

Quick Start

1. Download this repository.
2. Install — run install.bat, pick your GPU (CUDA 11.8 / 12.6 / 12.8 or CPU). It sets up portable Python, PyTorch and dependencies.
3. Run — run run.bat; the app opens in the browser, models download on first launch. Update with update.bat.

Or install one-click cross-platform via Pinokio — no install.bat needed.

Other Projects by @timoncool

Authors

• Nerual Dreming — Telegram | neuro-cartel.com | ArtGeneration.me
• Нейро-Софт — Telegram | portable AI builds

Acknowledgments

• Boson AI — Higgs Audio v3 — the TTS model
• multimodalart — transformers port of the model
• Slait/russia_voices — 743 Russian voice presets
• Moonshine ASR — reference auto-transcription
• pyloudnorm — EBU R128 loudness leveling · Gradio — UI framework

Support the Author

I build open-source software and do AI research. Most of what I create is free and available to everyone. Your donations help me keep creating without worrying about where the next meal comes from =)

All donation methods | dalink.to/nerual_dreming | boosty.to/neuro_art

• BTC: 1E7dHL22RpyhJGVpcvKdbyZgksSYkYeEBC
• ETH (ERC20): 0xb5db65adf478983186d4897ba92fe2c25c594a0c
• USDT (TRC20): TQST9Lp2TjK6FiVkn4fwfGUee7NmkxEE7C

Star History

License

The wrapper code is open. The Higgs Audio v3 weights are distributed by Boson AI under a Research & Non-Commercial license — this application is non-commercial. Voice cloning only with the consent of the voice owner; impersonation, fraud and any illegal use are prohibited. See the model card.