# FAQ Models are missing — what do I do? - Run `polyscribe --download-models` to pick and download models interactively. - For CI/non-interactive, set POLYSCRIBE_MODELS_DIR to a writable dir and run `polyscribe --update-models --no-interaction`. - You can also point to a specific file via `WHISPER_MODEL=/path/to/model.bin`. I get timeouts or slow downloads - Try again later or use a closer mirror (if available by setting upstream env vars or downloading manually into the models dir). - Ensure your network allows Hugging Face downloads. Non-interactive CI runs hang or fail - Add `--no-interaction` to disable prompts. - Set `POLYSCRIBE_MODELS_DIR` to a known location and pre-populate models or run `--update-models`. - Use `-q` to reduce noise in logs; use `-v` or `-vv` when debugging failures. GPU was not detected - Ensure you built with the matching feature (`gpu-cuda`, `gpu-hip`, or `gpu-vulkan`). - Install the relevant runtime (CUDA toolkit/driver, ROCm libraries, Vulkan loader/SDK) and ensure libraries are on the loader path. - Force CPU backend with `--gpu-backend cpu` to verify the rest of the pipeline. Which model directory is used in releases? - For packaged binaries, PolyScribe uses `$XDG_DATA_HOME/polyscribe/models` or `~/.local/share/polyscribe/models` by default. Override with `POLYSCRIBE_MODELS_DIR`. SRT timestamps look wrong - SRT times are derived from model timestamps. If your input has variable sample rate or corrupted timestamps, ensure ffmpeg can decode it; consider re-encoding the audio.