Question 1

How do I fix a Tauri blank or gray screen on Linux with NVIDIA and Wayland?

Accepted Answer

Start the development session with WebKitGTK's DMA-BUF renderer disabled and force the X11 backend:

```bash
WEBKIT_DISABLE_DMABUF_RENDERER=1 GDK_BACKEND=x11 just dev
```

For a compiled app, use:

```bash
WEBKIT_DISABLE_DMABUF_RENDERER=1 GDK_BACKEND=x11 ./your-app
```

`WEBKIT_DISABLE_DMABUF_RENDERER=1` works around the WebKitGTK/NVIDIA blank-rendering issue, while `GDK_BACKEND=x11` runs the window through X11/XWayland and avoids the related Wayland failure. You can add these variables to your Linux `justfile` recipe or shell profile until the upstream WebKitGTK/NVIDIA and Wayland issues are resolved. A separate `tao` 0.34.8 Wayland startup panic may require the patched `tao` version from PR #748.

Question 2

Why does Voicebox ignore VOICEBOX_DATA_DIR and keep using the C drive on Windows?

Accepted Answer

This is a confirmed bug in Voicebox 0.4.5: `VOICEBOX_DATA_DIR` was documented but not honored by either the backend or desktop app. The backend only processed `VOICEBOX_MODELS_DIR`, so the server still received the default `--data-dir` under `%APPDATA%\sh.voicebox.app`.

Fix: upgrade to a Voicebox release containing the PR "fix: honor VOICEBOX_DATA_DIR in both the backend and the desktop app", then restart the app after setting the environment variable.

Temporary workaround for 0.4.5: launch the backend explicitly with `--data-dir`, for example:

```powershell
cd "$env:APPDATA\sh.voicebox.app\backends\cuda"
.\voicebox-server-cuda.exe --data-dir "D:\VoiceboxData" --port 17493
```

Replace `D:\VoiceboxData` with your preferred location. Setting `VOICEBOX_DATA_DIR` alone will not work in 0.4.5.

Question 3

How do I re-enable microphone recording permission after accidentally clicking Deny?

Accepted Answer

Upgrade to the release containing the microphone-permission fix; when access is denied, the recording area now shows an inline prompt/notice to grant microphone access again. If you still do not see the prompt, reset the permission for the page: click the lock icon in the browser address bar → Site Settings → Microphone → set to Allow, then reload the page. Also ensure microphone access is enabled at the OS level: Windows Settings → Privacy & Security → Microphone, or macOS System Settings → Privacy & Security → Microphone, enable the browser/app, and restart it. Start recording again and the permission prompt should reappear.

Question 4

How to fix luxtts error "Could not load model openai/whisper-tiny" with AutoModelForCTC and AutoModelForSpeechSeq2Seq on Windows ARM64?

Accepted Answer

The `AutoModelForCTC` failure is expected because Whisper is a speech seq2seq model, not a CTC model. The important error is that `AutoModelForSpeechSeq2Seq` cannot find a PyTorch/Safetensors checkpoint. This usually means an outdated Transformers version, an incomplete offline cache, or a leaked `HF_HUB_OFFLINE` setting.

1. Update the packages used by luxtts:
```powershell
python -m pip install -U "transformers>=4.40" safetensors huggingface-hub
```
If luxtts uses a bundled Python environment, run this with that environment's `python`.

2. Remove the incomplete Whisper cache and force an online re-download:
```powershell
Remove-Item Env:HF_HUB_OFFLINE -ErrorAction SilentlyContinue
$env:HF_HUB_OFFLINE = "0"
Remove-Item -Recurse -Force "$env:USERPROFILE\.cache\huggingface\hub\models--openai--whisper-tiny"
```
Then restart luxtts and let it download `model.safetensors`.

3. Update luxtts/backend to a build containing the offline-mode fix `fix(offline): remove process-global offline guard from Qwen3 LLM loads`. The previous cached Qwen3 load could leave process-wide Hugging Face offline mode enabled, causing later model loads to use an incomplete cache.

As a last resort, if only TensorFlow weights are available and TensorFlow is installed, the model can be loaded with `from_tf=True`, but upgrading Transformers and downloading the Safetensors/PyTorch checkpoint is the recommended fix.

Question 5

How can I move the recording capture preview to another monitor or screen corner?

Accepted Answer

Use a release containing PR #796, which adds a persisted capture preview placement setting. Open Settings, choose the capture preview position from the six available placements, and start recording. The preview will appear on the monitor containing the cursor when recording starts instead of always appearing on the primary monitor. On older versions there is no built-in workaround; upgrade or build PR #796.

voicebox

Core Features

What It Can't Do

Use Cases

Tags

Getting Started

Similar Projects