# Transcription on CPU-only hosts

Telegram speech transcription runs locally, even when chat inference uses a
remote provider. With `OMNIUS_ASR_DEVICE` unset or set to `auto`, Omnius selects
CPU when hardware inventory establishes that CUDA hardware is absent. Missing
Torch, a driver failure, or a GPU visibility mask alone does not establish that
absence. Explicit CUDA selection remains CUDA.

Automatic CPU speech transcription uses a separate `.venv-whisper-cpu` runtime with
CPU-only Torch and preserves the selected Whisper model. It skips the automatic
transcribe-cli installation used by the CUDA path. Requested speaker diarization
continues through transcribe-cli rather than being silently discarded. Initial setup downloads the
runtime dependencies and requested model; later requests reuse them. Existing
explicit CPU overrides retain the `OMNIUS_ASR_ALLOW_CPU=1` requirement.

Transcription is a retryable observation: it does not modify its input or send a
Telegram message. Derived transcripts remain in host-owned artifact storage.
An interrupted transcription no longer creates an ambiguous external action.
When an older retained task resumes, only its authenticated, closed transcription
tickets can receive this classification repair. The repair supplies no transcript
or successful result. Unknown writes and uploads still require reconciliation.

Validation: 69 execution tests, 75 interruption tests, 7 retained Telegram cases,
2 actual Telegram tool-registration cases and 46 packaging checks passed. A real CPU-only Torch runtime decoded an English
speech clip using Whisper `base` in 8.9 seconds, leaving the input unchanged.
Hardware absence and driver-error distinctions were tested with deterministic
Linux, Windows and macOS inventories; the live CPU test explicitly selected CPU
on a host that also has GPUs. The smaller `tiny` model also decoded on CPU but
missed the stricter word-matching checks on the synthesized clip.
