The file is only the beginning
A transcription workflow decodes media, normalizes or resamples audio, detects speech, creates word or segment timings, assigns optional speaker labels, renders editable text, and exports it. A tool may succeed at one stage and fail at another. “Supports MP4” might mean audio can be extracted from some MP4 variants, not that every codec and damaged file will work.
Keep the original media unchanged. Work from a copy, record the tool and model version, and preserve enough timing information to revisit disputed passages. Long recordings should have a predictable resume or recovery path if the machine sleeps or the job is interrupted.
Local and cloud are processing choices, not quality guarantees
Local transcription can reduce the need to upload the recording and can keep working after required models are installed. It also uses the computer’s CPU, GPU, memory, storage, and power. Cloud transcription can centralize processing and collaboration, but requires examination of upload, retention, training, subprocessors, region, account, and deletion behavior.
A hybrid product can legitimately use both. Ask for a data-flow diagram or an operation-by-operation table rather than accepting a blanket label. Model download and software update traffic are different from uploading source audio, and the distinction should be explicit.
Speaker labels are hypotheses
Diarization divides speech into clusters such as Speaker 1 and Speaker 2. Overlap, short interjections, similar voices, room echo, remote-call processing, and background media can confuse those clusters. Speaker recognition that names a known individual is a separate and more sensitive capability.
Use neutral labels until a person checks them. Do not treat an automatically attributed statement as authoritative minutes, a legal admission, or a reliable identity record.
A transcript is complete after review
Review names, numbers, negations, specialist terms, action owners, dates, and passages with overlapping speech. Listen around the timestamp rather than guessing from surrounding text. Mark genuinely inaudible sections instead of inventing certainty.
Choose the export for the next task: plain text for search and drafting, timestamped text for verification, DOCX or PDF for circulation, JSON for software pipelines, and SRT or VTT for timed captions. Confirm encoding and line endings where another system will ingest the file.