A polished Windows WebUI for local Higgs Audio workflows: reference voice preparation, Higgs V2/V3 speech generation, Faster-Whisper transcription, dataset building, and LoRA training.
| Area | What it does |
|---|---|
| Higgs V3 TTS | Local Transformers inference with modern Higgs V3 speech generation. |
| Higgs V2 TTS | Local Higgs V2 generation for legacy/experimental workflows. |
| Reference Voice Library | Save reusable voice samples and transcripts under samples/. |
| LoRA Adapters | Load V2/V3 adapters from exp/ and switch them from the GUI. |
| Single Inference | Generate speech from target text with optional reference audio and seed control. |
| Dialogue Builder | Build multi-speaker turns, assign voice references, and concatenate the final WAV. |
| Long-Form / Batch | Split long text by paragraph, period, line, or speaker turns, then join chunks. |
| Faster-Whisper ASR | Transcribe reference samples or dataset audio locally. |
| Dataset Preparation | Convert curated audio folders into train/eval datasets under data/. |
| Training UI | Run single-speaker LoRA training for V2 and experimental V3 training. |
| Local Storage Layout | Models, cache, datasets, LoRAs, logs, samples, and outputs are organized by folder. |
Higgs Audio V3 TTS training is experimental and unofficial.
At the time this GUI was built, there is no official public Higgs V3 TTS fine-tuning guide or trainer equivalent to a mature supervised LoRA workflow. This app exposes an experimental V3 training path by adapting the available Higgs-style supervised workflow to the local V3 Transformers runtime.
The included V3 trainer uses teacher-forced cross entropy over delayed 8-codebook audio tokens. The following practical improvements were adapted from tuanh123789/Higgs-tts-3-finetune:
-
Audio-code cache: target/reference WAV files are encoded once and cached under
exp/<project>/audio_code_cache, so later epochs, eval passes, and resumed runs do not waste time re-tokenizing the same audio. -
V3 audio module persistence:
audio_embeddingandaudio_headare saved besideqwen3_loraashiggs_v3_audio_modules.ptand restored during inference/resume when present. This can make inference more faithful to the training/eval sound, because the adapter is loaded with the same V3 audio helper layers that were present during training instead of only the Qwen LoRA weights. It does not guarantee higher quality in every case, but it reduces one source of mismatch between training eval and later inference.
The GUI follows a complete local voice workflow:
Build a reusable reference voice library.
- Import or record short reference clips.
- Trim samples to a clean, stable range.
- Transcribe with Faster-Whisper or write transcripts manually.
- Save paired audio/transcript samples to
samples/.
Recommended reference clips are usually 5–30 seconds, clean, single-speaker, and free of background music or overlapping voices.
Generate speech with Higgs V2 or V3.
- Select the TTS model.
- Optionally load a LoRA adapter.
- Select a saved voice sample or upload reference audio.
- Tune the desired parameters under advanced settings.
- Save generated WAV files under
outputs/.
Long text should use chunking. Every chunk is a separate synthesis pass, so slight voice/prosody variation can happen between chunks.
Create multi-turn generated conversations.
- Add, copy, or delete speaker rows.
- Assign a different voice sample per turn.
- Generate each turn independently.
- Join the final dialogue with configurable silence gaps.
This is useful for podcasts, demos, character testing, and comparing voice references.
Prepare train/eval datasets from local audio folders.
- Supports common audio formats.
- Uses existing
.txtor.jsontranscripts when available. - Fills missing transcripts with Faster-Whisper.
- Writes prepared train/eval data under
data/. - Validates dataset structure before training.
Train or resume LoRA projects.
- Select a training project.
- Choose V2 or experimental V3.
- Configure batch size, gradient accumulation, learning rate, LoRA rank, bf16, logging, save, and eval intervals.
- Resume from discovered checkpoints.
- Store outputs, adapters, checkpoints, and eval audio under
exp/.
This project uses uv for fast and reproducible dependency management.
git clone https://github.com/Mixomo/Higgs_Audio_TTS_Easy_GUI.git
cd Higgs_Audio_TTS_Easy_GUI
Run from the project folder:
install.batThe installer will:
- Install
uvif needed. - Create the local virtual environment.
- Install non-Torch dependencies.
- Install the selected PyTorch backend last.
- Verify CUDA when a CUDA backend is selected.
| Option | Target |
|---|---|
| Auto-detect NVIDIA / CPU | Recommended first try. |
| CUDA 11.8 | NVIDIA GTX 10xx / Pascal compatibility. |
| CUDA 12.6 | NVIDIA RTX 20xx / 30xx. |
| CUDA 12.8 | NVIDIA RTX 40xx / 50xx. |
| CPU only | Compatibility mode; slow for TTS/training. |
| AMD DirectML | Experimental Windows AMD path. |
start.batThen open:
http://127.0.0.1:7860
- Windows 10/11.
- Modern web browser.
uvpackage manager.- NVIDIA driver compatible with the selected CUDA backend.
| Task | Minimum | Recommended |
|---|---|---|
| Faster-Whisper ASR | CPU or small GPU | 8 GB+ VRAM for large-v3 |
| Higgs V2 inference | 12 GB VRAM | 16 GB+ VRAM |
| Higgs V3 inference | 16 GB VRAM | 24 GB+ VRAM |
| LoRA training | 24 GB VRAM | 24 GB+ VRAM |
CPU mode is available for compatibility, but TTS inference and training are expected to be slow.
| Folder | Purpose |
|---|---|
models/ |
Final downloaded model folders. |
models/.cache/ |
uv, Hugging Face, Xet, Torch, temp, and runtime caches. |
samples/ |
Reference voice samples and transcripts. |
outputs/ |
Persisted generated WAV files. |
data/ |
Prepared training datasets. |
exp/ |
Training projects, LoRA adapters, checkpoints, and eval audio. |
logs/ |
App and training logs. |
config/ |
Local UI settings. |
.wav
.mp3
.flac
.ogg
.m4a
.txt
.json
For reference samples, use the same file stem:
my_voice.wav
my_voice.txt
- V3 can stop a generation around its internal end signal even when
max_new_tokensis high. - Use chunking for long text.
- Lower temperature and top-p values usually improve stability.
- Fixed seeds help compare settings; random seeds can improve natural variation.
RAS helps reduce audio-token loops, repeated syllables, repeated silences, or stuck generation patterns.
Typical starting point:
ras_win_len = 7
ras_win_max_num_repeat = 2
If generation loops, try lowering the repeat count. If speech becomes too clipped or unnatural, return to the default.
Torch compile is optional and can increase the synthesis speed by up to 2.3x (Nvidia GPU only)
- First generation can be slower (take up to 5 minutes) because kernels need warmup.
- Similar later generations can be significantly faster.
- Changing model, shape, reference pattern, LoRA adapter, or chunk layout can trigger new compile passes.
- Disable it if you prefer predictable speed over maximum throughput.
GUI Higgs V3 test on NVIDIA GeForce RTX 3090, same text, seed, and frame limit.
Max audio frames: 4096
Precision: auto -> bf16
Attention: SDPA
LoRA: none
Reference audio: none
| Mode | Load time | Frames | Audio length | Generation elapsed | RTF | Avg frames/s |
|---|---|---|---|---|---|---|
torch.compile off |
8.37s | 525 | 20.7s | 21s | 1.01 | 24.95 |
torch.compile on |
8.32s | 551 | 21.8s | 9s | 0.41 | 57.98 |
In this run, torch.compile was about 2.3× faster during frame generation once active. First-time compilation can still take up to 5 minutes to warmup.
| Symptom | Suggested action |
|---|---|
| CUDA OOM during inference | Lower max frames, unload ASR, disable compile, use chunking, or try lower precision if supported. |
| Voice changes between chunks | Enable shared chunk seed, use shorter chunks, or reuse the same reference sample consistently. |
| Output stops early | Use chunking; V3 may stop at its internal end signal. |
| First compile run seems frozen | Wait for kernel compilation; later runs with similar shapes should be faster. |
| ASR is slow | Use a smaller Faster-Whisper model or CPU-friendly settings. |
| Training overfits quickly | Lower learning rate, reduce max steps, increase eval frequency, and use cleaner/more varied data. |
Built for local use around:
- Boson AI Higgs Audio
- Train Higgs Audio repo
- tuanh123789/Higgs-tts-3-finetune
- Faster-Whisper / CTranslate2
GUI layout and workflow inspired by FranckyB Voice Clone Studio
This GUI is a practical local workflow tool, not an official Boson AI release. Higgs model licenses and usage restrictions still apply. Review the upstream model licenses before distributing generated audio, fine-tunes, or derived model artifacts.