| --- |
| license: other |
| license_name: boson-higgs-audio-v3-research-and-non-commercial-license |
| license_link: https://huggingface.co/bosonai/higgs-tts-3-4b/blob/main/LICENSE |
| language: |
| - af |
| - ar |
| - as |
| - ast |
| - az |
| - ba |
| - be |
| - bg |
| - bn |
| - bs |
| - ca |
| - ceb |
| - ckb |
| - cs |
| - cy |
| - da |
| - de |
| - el |
| - en |
| - eo |
| - es |
| - et |
| - eu |
| - fa |
| - fi |
| - fr |
| - ga |
| - gl |
| - gu |
| - ha |
| - he |
| - hi |
| - hr |
| - ht |
| - hu |
| - hy |
| - id |
| - is |
| - it |
| - jv |
| - ka |
| - kab |
| - kam |
| - kea |
| - kk |
| - kn |
| - ko |
| - ky |
| - la |
| - lb |
| - lg |
| - ln |
| - lt |
| - luo |
| - lv |
| - mhr |
| - mi |
| - mk |
| - ml |
| - mn |
| - mr |
| - ms |
| - mt |
| - ne |
| - nl |
| - no |
| - nso |
| - ny |
| - oc |
| - om |
| - pa |
| - pl |
| - ps |
| - pt |
| - ro |
| - ru |
| - rw |
| - sd |
| - sk |
| - sl |
| - sn |
| - so |
| - sq |
| - sr |
| - sv |
| - sw |
| - ta |
| - te |
| - tg |
| - tl |
| - tr |
| - ug |
| - uk |
| - umb |
| - ur |
| - uz |
| - vi |
| - xh |
| - zh |
| - zu |
| tags: |
| - text-to-speech |
| - speech-generation |
| - voice-cloning |
| - expressive-speech |
| - controllable-tts |
| - multilingual-tts |
| - cpp |
| - gguf |
| - windows |
| - whisper.cpp |
| - cuda |
| - higgs-audio-v3 |
| pipeline_tag: text-to-speech |
| --- |
| |
| # Higgs Audio v3 Studio Runtime Files |
|
|
| This repository hosts the downloadable runtime files for **Higgs Audio v3 Studio**, a Windows desktop app for local Higgs Audio v3 TTS, voice cloning, speech continuation, and multi-speaker generation. |
|
|
| GitHub app repository: https://github.com/Saganaki22/Higgs-Audio-v3-Studio |
|
|
| This repository is not the original upstream model release. It provides GGUF model builds, the Windows CUDA engine DLL package, checksums, and a manifest used by the desktop app downloader. |
|
|
| ## What This Is |
|
|
| Higgs Audio v3 Studio is a Rust/Tauri desktop application that runs a ported native C++/CUDA implementation of Higgs Audio v3 locally. |
|
|
| The app provides: |
|
|
| - Local TTS generation |
| - Voice clone workflow |
| - Continue speech workflow |
| - Multi-speaker workflow |
| - Speaker Gallery for reusable speaker identities |
| - WAV/MP3 export |
| - Local API server |
| - API streaming endpoint |
| - Whisper-assisted reference transcript workflow |
| - Model/engine download UI |
| - Hardware telemetry and VRAM diagnostics |
| - Engine dependency diagnostics for missing CUDA/MSVC runtime DLLs |
|
|
| ## App Download |
|
|
| Use the desktop app from GitHub releases: |
|
|
| https://github.com/Saganaki22/Higgs-Audio-v3-Studio/releases |
|
|
| Source code: |
|
|
| https://github.com/Saganaki22/Higgs-Audio-v3-Studio |
|
|
| ## Runtime Repository Layout |
|
|
| The app expects this Hugging Face repository layout: |
|
|
| ```text |
| / |
| ├─ manifest.json |
| ├─ checksums/ |
| │ └─ SHA256SUMS.txt |
| ├─ engines/ |
| │ ├─ audiocpp_engine.dll |
| │ ├─ cublas64_13.dll |
| │ ├─ cublasLt64_13.dll |
| │ ├─ MSVCP140.dll |
| │ ├─ VCOMP140.DLL |
| │ ├─ VCRUNTIME140.dll |
| │ └─ VCRUNTIME140_1.dll |
| └─ models/ |
| ├─ higgs-q4_k_m/ |
| │ └─ q4_k_m.gguf |
| ├─ higgs-q5_k/ |
| │ └─ q5_k.gguf |
| ├─ higgs-q6_k/ |
| │ └─ q6_k.gguf |
| ├─ higgs-q8_0/ |
| │ └─ q8_0.gguf |
| └─ higgs-bf16/ |
| └─ bf16.gguf |
| ``` |
|
|
| ## Available Downloads |
|
|
| | File | Purpose | |
| |---|---| |
| | `manifest.json` | App downloader manifest with file names, sizes, hashes, and recommended model metadata | |
| | `checksums/SHA256SUMS.txt` | SHA256 checksums for engine package and model files | |
| | `engines/audiocpp_engine.dll` | Windows CUDA engine DLL used by the Tauri desktop app | |
| | `engines/cublas64_13.dll` | NVIDIA CUDA 13 cuBLAS runtime DLL required by the engine | |
| | `engines/cublasLt64_13.dll` | NVIDIA CUDA 13 cuBLASLt runtime DLL required by cuBLAS | |
| | `engines/MSVCP140.dll` | Microsoft C++ runtime DLL | |
| | `engines/VCOMP140.DLL` | Microsoft OpenMP runtime DLL | |
| | `engines/VCRUNTIME140.dll` | Microsoft Visual C++ runtime DLL | |
| | `engines/VCRUNTIME140_1.dll` | Microsoft Visual C++ runtime DLL | |
| | `models/higgs-q4_k_m/q4_k_m.gguf` | Smaller quantized model | |
| | `models/higgs-q5_k/q5_k.gguf` | Balanced K-quant model | |
| | `models/higgs-q6_k/q6_k.gguf` | Higher-quality K-quant model | |
| | `models/higgs-q8_0/q8_0.gguf` | Recommended default model | |
| | `models/higgs-bf16/bf16.gguf` | Highest-fidelity BF16 model | |
|
|
| ## Recommended VRAM |
|
|
| | Model | Recommended VRAM | Notes | |
| |---|---:|---| |
| | `q4_k_m` | 8 GB | Smaller model for lower VRAM systems | |
| | `q5_k` | 9 GB | Balanced K-quant option | |
| | `q6_k` | 10 GB | Higher-quality K-quant option | |
| | `q8_0` | 12 GB | Recommended default quality/speed balance | |
| | `bf16` | 16 GB | Highest-fidelity testing model | |
|
|
| ## Engine Requirements |
|
|
| The prebuilt engine package is intended for: |
|
|
| - Windows x64 |
| - NVIDIA RTX 30xx, 40xx, or 50xx GPU |
| - CUDA 13 compatible NVIDIA driver |
| - Higgs Audio v3 Studio `0.2.31` or newer recommended |
|
|
| The `engines/` folder contains the app engine DLL plus the CUDA/MSVC runtime DLLs the current Windows engine build needs. |
|
|
| Important: |
|
|
| - `nvcuda.dll` is **not** included and should not be uploaded here. |
| - `nvcuda.dll` comes from the NVIDIA display driver. |
| - Users still need a working NVIDIA driver installed. |
| - Users should not need the full CUDA Toolkit installed if they use the app's `Download Engine DLLs` button. |
|
|
| The app can use either: |
|
|
| 1. DLLs already installed on the user's system, such as CUDA/MSVC runtime DLLs found through system paths. |
| 2. DLLs downloaded from this repository into the app's writable engine folder. |
|
|
| ## Engine Dependency Diagnostics |
|
|
| Higgs Audio v3 Studio checks the Windows loader dependencies before loading the engine. |
|
|
| The current validator checks for: |
|
|
| - `nvcuda.dll` |
| - `cublas64_13.dll` |
| - `cublasLt64_13.dll` |
| - `MSVCP140.dll` |
| - `VCOMP140.DLL` |
| - `VCRUNTIME140.dll` |
| - `VCRUNTIME140_1.dll` |
|
|
| If a runtime DLL is missing, users can press `Download Engine DLLs` in the app to download the files from this repository. |
|
|
| If `nvcuda.dll` is missing, users need to install or update their NVIDIA driver. |
|
|
| ## C++ Port Overview |
|
|
| The desktop app uses a ported native C++/CUDA inference engine instead of running the original Python pipeline directly. |
|
|
| At a high level: |
|
|
| - Higgs Audio v3 model weights are loaded from GGUF files. |
| - The native engine exposes a C ABI through `audiocpp_engine.dll`. |
| - The Rust/Tauri backend dynamically loads the DLL and calls the native API. |
| - The TypeScript frontend talks to Rust through Tauri IPC. |
| - The same engine path is used by the UI and the local API server. |
| - Whisper.cpp support is integrated into the engine path for reference transcript assistance. |
|
|
| Main runtime layers: |
|
|
| ```text |
| Tauri UI |
| ↓ |
| Rust backend / local API / queue |
| ↓ |
| C ABI engine wrapper |
| ↓ |
| Ported C++ Higgs Audio v3 runtime |
| ↓ |
| ggml / CUDA backend |
| ↓ |
| GGUF Higgs model files |
| ``` |
|
|
| ## Runtime Optimisations |
|
|
| The current app and engine include several production-focused improvements: |
|
|
| - Live streaming playback for generated audio |
| - API streaming through newline-delimited JSON events |
| - Unified UI/API generation queue |
| - Cancel checks inside the native decode loop |
| - Runtime graph cleanup after generation, cancellation, and error paths |
| - Per-stage VRAM diagnostics in the app Command Centre |
| - F16 decode KV cache enabled by default in the CUDA engine |
| - F32 decode KV fallback available for diagnostics |
| - Unsupported mixed K/V cache modes disabled to avoid CUDA flash-attention fatal paths |
| - Saved speaker reference cache support through `.hspkcache` |
| - Speaker ZIP import/export for reusable speaker identities |
| - Manifest/checksum-based model and engine downloads |
| - Engine dependency preflight for clearer Windows DLL loading errors |
|
|
| ## Quantization Notes |
|
|
| The GGUF builds in this repository are provided so users can choose a quality/VRAM tradeoff from inside the app. |
|
|
| - `q4_k_m`: smallest supported Higgs model option in this repo |
| - `q5_k`: middle option between Q4 and Q6 |
| - `q6_k`: higher-quality K-quant option below Q8 |
| - `q8_0`: recommended default for most users with enough VRAM |
| - `bf16`: highest-fidelity build, largest VRAM requirement |
|
|
| The app keeps the small model assets/config files bundled with the installer/portable package where possible, while the large GGUF weights are downloaded separately from this Hugging Face repository. |
|
|
| ## API Support |
|
|
| Higgs Audio v3 Studio includes a local API server. |
|
|
| Supported API capabilities include: |
|
|
| - Plain TTS |
| - Saved-speaker voice clone |
| - Continue speech |
| - WAV output |
| - MP3 output |
| - Streaming NDJSON audio/progress events |
| - Saved speaker discovery |
| - API Command Centre logs in the desktop app |
|
|
| For API examples and current endpoint details, see the GitHub README: |
|
|
| https://github.com/Saganaki22/Higgs-Audio-v3-Studio#http-api-streaming-and-command-centre |
|
|
| ## Installation |
|
|
| Recommended path: |
|
|
| 1. Download the latest installer or portable build from GitHub releases. |
| 2. Open Higgs Audio v3 Studio. |
| 3. If the engine or runtime DLLs are missing, click `Download Engine DLLs`. |
| 4. Download or select a Higgs GGUF model. |
| 5. Load engine. |
| 6. Load model. |
| 7. Generate audio. |
|
|
| GitHub releases: |
|
|
| https://github.com/Saganaki22/Higgs-Audio-v3-Studio/releases |
|
|
| ## Manual Engine Folder |
|
|
| If placing files manually, keep the engine package together: |
|
|
| ```text |
| engines/ |
| ├─ audiocpp_engine.dll |
| ├─ cublas64_13.dll |
| ├─ cublasLt64_13.dll |
| ├─ MSVCP140.dll |
| ├─ VCOMP140.DLL |
| ├─ VCRUNTIME140.dll |
| └─ VCRUNTIME140_1.dll |
| ``` |
|
|
| Do not upload or redistribute `nvcuda.dll`. It belongs to the NVIDIA driver. |
|
|
| ## Checksums |
|
|
| Checksums are provided in: |
|
|
| ```text |
| checksums/SHA256SUMS.txt |
| ``` |
|
|
| The desktop app and users can use this file to verify downloaded runtime/model files. |
|
|
| ## Safety And Responsible Use |
|
|
| Do not use Higgs Audio v3 Studio, Higgs Audio v3, or any voice cloning workflow to impersonate people without consent, create deceptive or malicious voices, defraud people, bypass identity checks, harass others, or cause harm. |
|
|
| Only generate speech when you have the rights and consent required for the source voice, transcript, and intended output use. |
|
|
| ## Credits |
|
|
| This project depends on the upstream Higgs Audio v3 model work by Boson AI. |
|
|
| Upstream model: |
|
|
| https://huggingface.co/bosonai/higgs-audio-v3-tts-4b |
|
|
| Desktop app and ported runtime repository: |
|
|
| https://github.com/Saganaki22/Higgs-Audio-v3-Studio |
|
|
| Whisper.cpp: |
|
|
| https://github.com/ggml-org/whisper.cpp |
|
|
| NVIDIA CUDA runtime components are provided under NVIDIA's CUDA Toolkit license terms. |
|
|
| Microsoft Visual C++ runtime components are provided under Microsoft's Visual Studio / Visual C++ Redistributable license terms. |
|
|
| ## Citation |
|
|
| ```bibtex |
| @misc{bosonai_higgs_audio_tts_v3_2026, |
| title = {Higgs TTS 3: Conversational Speech for Voice AI from Boson AI}, |
| author = {Boson AI}, |
| year = {2026}, |
| howpublished = {https://huggingface.co/bosonai/higgs-tts-3-4b}, |
| } |
| ``` |
|
|
| ## License |
|
|
| The Higgs Audio v3 / Higgs TTS 3 model weights and upstream model assets are governed by the **Boson Higgs TTS 3 Research and Non-Commercial License**. |
|
|
| The desktop app and native port code are maintained separately on GitHub: |
|
|
| https://github.com/Saganaki22/Higgs-Audio-v3-Studio |
|
|
| Check all upstream licenses before redistributing weights, using generated audio commercially, or packaging this project into another product. |
|
|
