Windows + NVIDIA users hit `Could not locate cudnn_ops_infer64_8.dll`:
WhisperX/faster-whisper run on CTranslate2, which needs cuDNN 8, but PyTorch
2.8 ships cuDNN 9 and the side-loaded `cudnn8_compat` libs were missing from
the venv. The PyTorch-Whisper backend should have been the fallback, but it
errored "set OMNIVOICE_PRELOAD_TTS_ASR=1" because it only worked when the TTS
model preloaded an ASR head.
- `PyTorchWhisperBackend._ensure_pipe()` now builds its OWN transformers ASR
pipeline on demand (PyTorch stack → cuDNN 9, no CTranslate2/cuDNN-8), without
loading the full TTS model and without the preload env var. A constructor-
passed pipe (when the TTS model already has one) is still reused. Model is
overridable via OMNIVOICE_PYTORCH_ASR_MODEL.
- dub_core transcribe preflight no longer hard-rejects pytorch-whisper when no
pipe is preloaded — it lazy-loads; any failure surfaces per-chunk with the
real cause.
So a Windows box without cuDNN 8 can switch ASR backend to "PyTorch Whisper"
in Settings → Models and transcription works. Docs: troubleshooting entry.
Tests: tests/test_pytorch_whisper_fallback.py (lazy standalone build, reuse of
a passed pipe, no get_model() call, env override). Full tests/ suite: 700 pass.
Does NOT close#255 — pending the reporter confirming the fallback works on
their machine; the cuDNN-8 install gap (faster-whisper path) is a follow-up.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
