Whisper workflows hit Apple Silicon limits

This is really a document reconstruction problem disguised as a transcription question: accuracy, instruction following, and error correction matter more than flashy generative output.

• –The workflow mixes OCR cleanup, ASR cleanup, de-duplication of small talk, and strict report writing, so a strong speech model alone is not enough.
• –Inputs in the 25-50k character range make long-context reliability and unified-memory limits on laptop-class Apple Silicon a real constraint.
• –German grammar, zero-summarization requirements, and “don’t miss anything” rules push the task toward deterministic, low-temperature models with strong obedience rather than creative writing.
• –The hardware question matters because local users want enough RAM headroom to run long contexts comfortably without jumping straight to a desktop workstation.