Warning
This project is no longer developed. Text-to-speech on Higgs Audio v3, voice cloning, the AI text director, podcasts and audiobooks now live in Speech Studio. It brings together the latest AI technology: text-to-speech with Higgs Audio v3, Fish Audio Drama 3, Qwen3-TTS and Supertonic 3, voice cloning, voice design from a description, audiobooks and podcasts, speech recognition with speaker separation on Parakeet, Nemotron and CrisperWhisper, and voice conversations with an AI agent. Free and offline, on Windows, macOS and Linux.
Portable local text-to-speech built on Higgs Audio v3 TTS — expressive speech in 100+ languages, zero-shot voice cloning, an AI text director, plus Podcast and Audiobook modes. 100% offline, one click.
🚀 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.
- 🎙️ 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.
| OS | GPU | Status | Acceleration |
|---|---|---|---|
| Windows 10/11 | NVIDIA RTX 30xx–50xx | ✅ tested | CUDA 12.8 + Triton (torch.compile ~2×) |
| Windows 10/11 | NVIDIA RTX 20xx | ✅ expected | CUDA 12.8 + Triton |
| Linux x64 | NVIDIA RTX 20xx–50xx | ✅ expected | CUDA 12.8 + Triton |
| Linux aarch64 | NVIDIA DGX Spark / Jetson | ✅ expected | CUDA 13.0 |
| Windows | AMD RDNA3+ | ✅ expected | DirectML |
| Linux | AMD RDNA3+ | ✅ expected | ROCm 6.3 |
| macOS | Apple Silicon M1–M4 | ✅ expected | MPS |
| macOS | Intel | torch CPU | |
| Any | CPU only | CPU |
Higgs uses PyTorch SDPA (flash kernels built in) and does not need external Flash-Attention 2. The local
install.batbuild targets NVIDIA Windows; full cross-platform support is via Pinokio.
| VRAM | TTS mode | LLM director |
|---|---|---|
| 24 GB+ | bf16 (~11 GB) | 9–12B in 4-bit (~6–8 GB) |
| 12 GB | 8-bit (~6–7 GB) | 4–9B in 4-bit (~3–6 GB) |
| 6–8 GB | 4-bit (~3.5 GB) | 2–4B in 4-bit (~1.5–3 GB) |
| CPU | works, very slow | — |
Models (~9 GB TTS + LLM) download automatically on first run.
- Download this repository.
- Install — run
install.bat, pick your GPU (CUDA 11.8 / 12.6 / 12.8 or CPU). It sets up portable Python, PyTorch and dependencies. - Run — run
run.bat; the app opens in the browser, models download on first launch. Update withupdate.bat.
Or install one-click cross-platform via Pinokio — no install.bat needed.
Other Projects by @timoncool
| Project | Description |
|---|---|
| VoxCPM2 Portable | Multilingual TTS + Voice Design + LoRA fine-tuning |
| Qwen3-TTS | Portable text-to-speech with voice cloning |
| ACE-Step Studio | AI music studio — songs, vocals, covers, videos |
| Foundation Music Lab | Music generation + timeline editor |
| VibeVoice ASR | Portable speech recognition |
| LavaSR | Portable audio enhancement |
| SuperCaption Qwen3-VL | Portable image captioning |
| VideoSOS | AI video production in the browser |
- Nerual Dreming — Telegram | neuro-cartel.com | ArtGeneration.me
- Нейро-Софт — Telegram | portable AI builds
- 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
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
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.