// docs / troubleshooting
Troubleshooting
Root-cause fixes for the most common setup issues — capture, CUDA, the WebSocket connection, and dependency errors — followed by the local-first diagnostics export.
Captions are not appearing
Verify the selected audio device in the panel. If you are capturing desktop sound on Windows, make sure you enabled the correct WASAPI loopback channel.
CUDA error / out of memory
LiveAudio monitors VRAM before each transcription. When free memory drops below about 500 MB it frees the CUDA cache automatically and surfaces “GPU saturada - VRAM baja”. If it still fails: switch the inference device from cuda to cpu, or load a smaller model (tiny or base), and close other VRAM-heavy programs (OBS with NVENC, games). Updating your NVIDIA driver resolves most CUDA runtime failures.
“cublas64_12.dll not found” (CUDA)
This is a DLL-resolution issue, not a missing install. The cu121 extra routes torch/torchaudio from download.pytorch.org/whl/cu121 (and cpu from /whl/cpu); exactly one extra must be selected. On Windows, ctranslate2 — faster-whisper’s backend — loads cublas64_12.dll and cuDNN 9 via the system PATH, not Python’s DLL directories, so LiveAudio prepends torch/lib to PATH at startup (liveaudio/utils/dllpath.py). If you run from a custom environment and hit this error, make sure the CUDA torch wheels are installed and that torch/lib is reachable. The cu121 stack pins torch>=2.4,<2.6 (no 2.6 wheels are published there, and ≥ 2.4 ships the cuDNN 9 that ctranslate2 links against). Driver floor is 525.
OBS does not connect
Make sure the engine is running — port 8765 must be listening. If your firewall or antivirus blocks local traffic, add a rule to allow localhost connections on the WebSocket port.
No module named 'torch'
Installer: run the launcher with --reinstall. Developers: run uv sync --extra cpu (or --extra cu121), then uv run liveaudio.
Audio cuts out or long silences
Increase the silence detection value (around 1.0–1.5 s) so heavily paused phrases are not cut mid-sentence.
Windows SmartScreen warning
Install & download recovery
Issues during the first-run bootstrap (or an interrupted CUDA download) are recoverable without reinstalling from scratch:
| Problem | What to do |
|---|---|
| Interrupted download (big CUDA install) | Just re-run the launcher. uv caches per-wheel, so completed wheels are not re-downloaded; the source zip is re-fetched and checksum-verified. |
| “SHA256 mismatch” | The download was corrupted or tampered with. The launcher retries automatically (3 attempts with backoff). Persistent failures usually mean a proxy or antivirus is rewriting downloads; verify against SHA256SUMS.txt. |
| Corrupted / half-broken install | Run the launcher with --reinstall — it wipes app/ and .venv, keeps your device preference, and bootstraps fresh. |
| Wrong backend installed | Re-run with --device cpu or --device cuda; the launcher re-syncs and persists the choice. |
| Where are the logs? | bootstrap.log in the install root (%LOCALAPPDATA%\LiveAudio on Windows, ~/.local/share/liveaudio on Linux, or data/ next to the launcher in portable mode). The error dialog also has an “Open log” button. |
Local diagnostics
LiveAudio ships local-first diagnostics — no telemetry is sent anywhere. Use them when OBS stops receiving subtitles, the ASR falls behind, or audio reconnects too often.
- How: reproduce the issue, click Export diagnostics in the app, then review the exported JSON (process states, queue sizes, audio/asr/ws states).
- Privacy: the export never includes raw audio or full transcripts, and it sanitizes secrets, tokens and sensitive paths.
| Key | Value |
|---|---|
diagnostics_enabled | true / false |
diagnostics_level | minimal / deep |
diagnostics_export_dir | path or null |