AudioEditor

Cleans recorded audio automatically — strips filler words, false starts, stutters, and dead air, attenuates breaths, and crossfades every cut. It transcribes the file at the word level, has Claude classify each segment (KEEP, CUT_FILLER, CUT_FALSE_START, CUT_STUTTER, CUT_DEAD_AIR), then executes the cuts with ffmpeg. An optional Cleanvoice pass adds final polish. Modes: --preview, --aggressive, --polish.

Cleaning a recording by hand means scrubbing a waveform for every "um," half-started sentence, and three-second silence, then crossfading each cut so it doesn't click. It's slow and tedious, and a blunt auto-tool over-cuts — it kills the rhetorical pause along with the accidental one, or leaves an audible seam where it spliced. This pipeline tells deliberate pauses apart from dead air, fills gaps with room tone, and crossfades each edit, so the output sounds clean rather than chopped.

Whisper produces word-level timestamps, Claude classifies each segment (distinguishing rhetorical emphasis from accidental repetition), and ffmpeg executes the cuts with 40ms qsin crossfades, room-tone gap fill, and breath attenuation at 50% volume rather than removal. An optional Cleanvoice API pass handles mouth-sound removal, residual filler, and loudness normalization.

Pipeline

Audio Input
 |
[Transcribe] Whisper word-level timestamps (insanely-fast-whisper on MPS)
 |
[Analyze] Claude classifies each segment:
 | KEEP / CUT_FILLER / CUT_FALSE_START / CUT_EDIT_MARKER / CUT_STUTTER / CUT_DEAD_AIR
 | Distinguishes rhetorical emphasis from accidental repetition
 |
[Edit] ffmpeg executes cuts:
 | - 40ms qsin crossfades at every edit point
 | - Room tone extraction and gap filling
 | - Breath attenuation (50% volume, not removal)
 |
[Polish] (optional) Cleanvoice API final pass:
 - Mouth sound removal
 - Remaining filler detection
 - Loudness normalization

Output: cleaned MP3/WAV

Capture-chain noise gates (recorder filters, macOS Voice Isolation) truncate audio to digital zero with no fade. Every gate close/open edge is a step discontinuity; leveling (+gain/compression) amplifies each edge into an audible tick — 774 edges in one 13-minute recording shipped in two public launch videos and drew listener complaints.

The gate, non-negotiable for any audio that will be leveled or published:

1. Before ANY leveling/gain/compression: GateScan the raw source. Gate activity present → run GateRepair FIRST (edges smoothed at low level), THEN level. Boosting first turns inaudible edges into loud ticks.
2. Before ANY upload/publish: GateScan the FINAL ENCODED file must exit 0. AAC encoding re-introduces steps near silence — scan the encode, not the intermediate WAV (GateRepair --finalize owns this loop).

Example 1: Clean a podcast recording

User: "clean up the audio on this podcast file"
-> Invokes Clean workflow
-> Runs full pipeline: transcribe -> analyze -> edit
-> Outputs cleaned MP3 with filler words, stutters, and dead air removed

Example 2: Preview edits before applying

User: "show me what edits you'd make to this recording"
-> Invokes Clean workflow with --preview flag
-> Transcribes and analyzes, shows proposed edits without modifying audio
-> User reviews edit list, then runs again to apply

Example 3: Aggressive clean with cloud polish

User: "aggressively clean this audio and polish it"
-> Invokes Clean workflow with --aggressive --polish flags
-> Tighter thresholds for filler detection
-> Cleanvoice API pass for mouth sounds and normalization

• Transcription accuracy varies with audio quality. Background noise, multiple speakers, and accents reduce accuracy.
• Cut detection is heuristic-based. Always preview edits before committing — automated cuts can remove intentional pauses.
• Cloud polish uploads audio to external service. Confirm the user is okay with cloud processing for sensitive content.
• dB-domain cliff detectors false-positive on fade feet. A legitimate cosine fade into digital zero has infinite dB slope at its foot, so any >NdB/2ms detector flags it forever. Verify repairs with sample-domain STEP detection (GateScan), never dB slopes. (2026-07-13)
• One-sided iterative fades don't converge on stray boundary impulses. A single sample sitting on a gate boundary survives repeated one-sided fades; fade-down-then-up "mutes" preserve a blip's edges and its leading step. The V-notch (cosine to zero at the boundary, both sides) and hard-zeroing inside gated silence are the converging fixes. (2026-07-13)
• Meter-clean ≠ ear-clean. Scanners find ticks; repairs can create audible holes the tick-scanner calls clean (the 2026-07-12 duck incident). Match the probe to the defect class a human hears, keep repairs minimal-touch, and give the principal before/after listen clips at swap time.