Define the transcript before changing the recording
This workflow serves journalists, research interviewers and editors handing off one recorded conversation. Agree whether the result is verbatim, lightly edited or a reading copy. Do not silently remove hesitations, qualifiers or negation from material intended for quotation. Identify who can approve names and who owns final review. Automatic speaker identification, translation, authenticity assessment and legal admissibility are outside this page.
Keep the original locally, record its SHA-256 digest and inspect its duration and channels. A hash identifies a file version; it does not establish who recorded it or whether its contents are true. Separate microphone channels can contain different speakers, so a multichannel warning is not an instruction to collapse them blindly. If extraction or channel selection is needed, perform it in an approved media process and record the relationship to the original.
Diagnose first and keep cleanup optional
Use the waveform to locate suspicious levels and the spectrogram to inspect persistent noise, then listen at those times. Neither image reveals the words. The dictation checker estimates readiness from format, activity, clipping and noise separation. Without a suitable non-speech reference, noise separation can be unavailable; never rewrite that missing measurement as a clean pass.
Create a cleanup preview only for a real obstacle. Save the assistant's suggested filter chain and chosen output settings. Its diagnostic analysis uses mono PCM, so reports are not separate speaker-channel assessments. Compare source and derivative at comparable playback level, especially quiet replies, consonants and simultaneous speech. More pleasant sound or a higher score is not proof of better transcription. Keep an unprocessed input when cleanup loses information; substantial restoration belongs in the related repair workflow.
Treat transcription as a draft with a known clock
The transcriber sends audio to an external AI provider. Confirm this is authorized before use; an authorized local copy is not automatically permission for external processing. Check the applicable duration limit instead of assuming all long interviews fit. For separately prepared excerpts, keep a table of original start offsets and restore global time references manually; this Hub does not perform automatic long-interview assembly.
Choose text for reading, or SRT, VTT or JSON for review with time references. Save the returned timing provenance. Provider timestamps may be replaced by silence-aligned approximate timing when unavailable. Replay the entire draft against the original, paying particular attention to chunk joins, repeated or missing words, names, dates, amounts and negation. Assign speaker labels from verified context, not vocal guesses. Mark uncertain speech with an original-time range and an explicit uncertainty label rather than inventing a completion.
Approve the words and the exact handoff files
For SRT or VTT, the sync checker compares cue timing with likely speech activity. It does not check wording, identify speakers or automatically repair drift. A cue can overlap speech while containing the wrong words. Review opening, middle and closing anchors and every flagged interval by playback; plain text and JSON references require manual checks. Caption readability and publication delivery belong in the accessibility workflow.
Open the exact final text or timed file in the receiving application, check encoding and line breaks, and verify that its transcript policy and time origin are stated. Deliver the approved draft revision, source identifier, derivative identifiers where used, processing settings, reports and uncertainty list with reviewer approval. Hash the final files for version tracking, but do not expect them to match the source hash. Store the package outside temporary processing links. A traceable handoff makes limits visible rather than hiding them behind an AI success message.