Group recordings by conditions, not filenames
This workflow serves interview editors, course-recording teams and oral-history organizers who need consistent cleanup copies from several related recordings. It does not transcribe speech, deduplicate a music library or authenticate evidence. Retain originals, record SHA-256 and inspect codec, duration and channels. A hash identifies a file version, not recording truth or processing authorization. Review confidentiality and permission before upload.
Create a small ledger linking source, group, microphone, gain, noise conditions and reviewer. A room label or matching codec does not establish a common noise floor. Compare recordings from the beginning and end of the session and after device or background changes. Treat separate microphones and channels deliberately; do not discard speaker separation merely to simplify processing. General damage repair belongs in the related restoration workflow.
Save a noise reference that actually represents the problem
Listen while using waveform and spectrogram to locate a genuine noise-only interval. Quiet words, breaths that must remain, music and another speaker are not unwanted noise by default. The tool accepts a selected interval of 0.5 to 30 seconds and rejects near digital silence; ensure the actual in-file interval is adequate rather than relying on typed bounds alone.
Use extract-only to retain a JSON profile and note the original file and time range. The profile contains measurements such as band levels, flatness, hum evidence and recommended settings, but reuse is not full-spectrum template subtraction. The implemented chain uses the recommended noise-floor value and optional high-pass, while the separately chosen strength sets reduction. Save the JSON from a trusted extraction; passing the import shape check does not prove compatibility or safety. Do not manually edit it to make a poor match look acceptable.
Pilot branches without accumulating processing
Apply the saved profile to representative originals, beginning conservatively. Compare a separate speech-enhancer candidate only where useful, also starting from the original. The enhancer has its own documented chain; it does not import the noise-profile JSON. Treat these as alternatives, not a compulsory serial chain. If neither preserves speech, retain the original and escalate the defect rather than repeatedly processing an already cleaned copy.
Compare at similar playback level so a louder candidate does not win automatically. Listen to consonants, low-level replies, simultaneous voices, pauses and changes in background noise. Clarity scores are heuristics, not STI measurements or transcript accuracy. Loudness reports help describe level changes but cannot verify words. Before/after floor estimates from quiet blocks may reflect wanted speech or gain changes, and missing readings remain unknown. Record why the chosen candidate helps and which artifacts make another unacceptable.
Review every reused output and verify the handoff
Reuse means applying the stored profile in separate runs, not an automatic batch job or a guarantee that later files match. Keep each reported chain, strength, source revision and profile revision in the ledger. Recheck representative passages in every output, including the quietest wanted voice; spot-checking one file cannot approve the others. Split the group and extract a new profile when conditions change. An accepted remaining noise bed is preferable to words erased by stronger reduction.
Probe exact delivered files and confirm their expected duration, channel layout and codec. Processing is not guaranteed sample-identical; explain material discrepancies rather than assuming they are harmless. Hash each final audio file independently; its digest should not match a transformed source. Deliver locally retained originals, approved copies, profile JSON, reports, per-file decisions and human sign-off, without relying on expiring output links. State unresolved defects and rejected files. Continue to the transcription workflow only after cleanup approval; a documented audio handoff is not a verified transcript or a forensic certification.