Should I vibe code
Remove filler words, mouth sounds, and long silences from a transcript-aligned edit list
Cutting the 'ums' is one afternoon. Making the cut inaudible is the entire product.
?
Their verdict, the Subscription — 10 hours/month price and the build-time estimate come from their entry, MIT-licensed. Checked 2026-08-04.
?
Our verdict, the regret score and everything below it. Editorial and unsponsored — nobody can pay to be moved.
The honest answer
why the verdict is what it is
Whisper hands you word-level timestamps and ffmpeg cuts on them, so the first version really is an afternoon — auto-editor has been doing the silence half in public for years. The demo is genuinely thrilling: ten minutes of tape, three hundred 'um's gone, and you start wondering what the subscription was ever for. The subscription is for the next ten minutes. Whisper's word boundaries drift by tens of milliseconds and drift differently around plosives, so cuts that measure correct sound wrong — clipped consonants, breaths that stop mid-exhale, room tone that changes between one sentence and the next. Multitrack is where the gap becomes a canyon, because the actual trick here is aligning three guests' separate files so that removing Alice's filler doesn't leave Bob's laugh hanging over a hole. Nobody gets hurt. Your show just quietly starts sounding slightly synthetic, and you will be the last person to notice, because you have heard the file forty times.
What actually breaks
not "if". the specific failures.
- Word timestamps that run 40ms optimistic, so every cut takes the front off the next consonant and the whole episode develops a lisp
- Breath and room tone removed along with the silence, leaving speech that sounds vacuum-packed in a way listeners feel rather than identify
- A destructive render over the only copy of a two-hour interview with someone who will not sit down with you twice
- Multitrack drift, where files start at different offsets and an edit list computed on one track lands mid-word on another
- A model or library update that changes filler detection, so episode 41 sounds like a different show from episode 40
- Memory, the first time you feed it a real hour of 48kHz stereo instead of the two-minute sample you developed against
- Guest audio quietly uploaded to a hosted transcription API you never mentioned when you asked them to record locally
Is that you?
the verdict is a default, not a law
- You are trimming silence and dead air only, which is the solved half and already has good tools
- Inputs are never modified and every run writes a new file next to an inspectable edit list
- It is your own voice on your own show and you listen to the render before it goes out
- Transcription runs locally, so nobody's raw audio leaves the machine it was recorded on
- The pipeline renders in place, or the master exists only on the machine doing the rendering
- You are editing other people's episodes for money, because their voice is the thing you are being paid to protect
- Guests were told the recording stays with you and your first pipeline step posts it to an API
- The show is multitrack and you were planning to treat each track independently
If you build it anyway
the checklist, then the prompt that enforces it
- Never write over an input. Every run produces a new file plus an edit decision list, and refuses to start if the output path already exists.
- Emit the EDL before the audio — CSV plus an Audacity or Reaper marker file. A cut list you can read beats a render you have to trust.
- Crossfade every boundary, 20–30ms per side. A sample-accurate cut on a plosive is an audible click, and clicks are how homemade editing announces itself.
- Shorten pauses rather than deleting them, and keep real room tone in the gap. Digital silence between sentences is the single most recognisable artefact of automated cleanup.
- Put a floor under filler detection: skip candidates under about 120ms, skip anything whose neighbours sit closer together than the crossfade, and log every skip.
- For multitrack, prove alignment with a measured cross-correlation offset per file before applying a single cut, then apply cuts to the timeline rather than to one track.
- If transcription goes to a hosted API, say so in the README, read the provider's retention policy, and default anything with a guest on it to a local model.
- Keep a spot-check script that plays ten seconds either side of a random cut, original against processed, and run it before every publish.
I am building an audio cleanup pipeline for a podcast — filler-word removal,
silence trimming, light noise reduction. My recordings are irreplaceable. Build
like you believe that, and push back when I ask for shortcuts.
1. Never modify or overwrite an input file. Every run writes to a new output
path and refuses to start if that path already exists.
2. Before cutting anything, produce a reviewable edit decision list: start and
end timestamps, the reason for each cut, and the transcript span. Emit it as
CSV and as an Audacity/Reaper marker file. I must be able to reject a run
without having listened to it.
3. Only then render audio, and render from the EDL, so the list and the file
can never disagree.
4. Crossfade every boundary, 20-30ms per side. A sample-accurate cut on a
plosive is an audible click, and a click is what gives homemade editing away.
5. Do not replace pauses with digital silence. Shorten them and keep real room
tone in the gap, because speech with no room tone sounds vacuum-packed.
6. Put a floor under filler detection: skip candidates shorter than ~120ms,
skip any whose neighbouring words are closer than the crossfade, and log
every skip with its timestamp.
7. For multitrack, align first and prove it with a measured cross-correlation
offset per file. Apply cuts to the timeline, never to one track alone.
8. Stream through ffmpeg. Do not load an hour of 48kHz stereo into memory
because it worked on the two-minute sample.
9. If transcription goes to a hosted API, say so loudly in the README, name the
provider's retention policy, and make a local model the default for anything
with a guest on it.
10. Give me a spot-check script that plays ten seconds either side of a random
cut, original against processed, so I hear artefacts before listeners do.
11. Out of scope: music beds, loudness mastering, and anything claiming to fix
a bad room. If what I want is a de-fillering model that already works, say
so and tell me ten hours of Cleanvoice is $11.That one keeps you out of trouble. For the prompt that actually builds it, canivibecodeit.com has one.
their build prompt ↗Or don’t build it
the boring option, and the way back out
You publish weekly and the show is the point. Ten hours a month is $11, which is less than the time you will spend tuning crossfade lengths in month two, and their filler model has been trained on far worse audio than yours. Build the silence trimmer for fun — that part genuinely works. Rent the model that decides what counts as an 'um'.
$11/mo is cheaper than your weekend.
There is barely one to plan, and that is the point: keep originals untouched and the tool stays a filter you can simply stop running. Store the raw multitrack for every episode in a folder you back up, keep the EDLs beside them, and any future editor — human or hosted — starts from exactly where you did. The only version with a real exit problem is the one that renders in place, because there the exit is a re-record.
Actively developed CLI that cuts silence and dead air from audio and video, and does the easy half of this properly.
The transcription model everything in this category sits on, including the word-level timestamps your cuts will depend on.
Long-running open-source multitrack audio editor, and the thing your edit list should be able to open in.
Questions
auto-editor already exists and it is free. Why is this not a ship it?
Because auto-editor solves silence, and silence is the deterministic half. A gap is a gap; you can measure it. Filler words are a classification problem with a fuzzy boundary, mouth sounds are a signal problem, and both have to land on a cut point that sounds natural to a human ear. If silence trimming is genuinely all you need, take auto-editor and skip the subscription — that part is a ship it.
What is the failure that actually costs me something?
A pipeline that renders over its input. Every other problem here is cosmetic and fixable by re-running with different parameters, because the original still exists. Overwrite the master of an interview and the fix is asking a stranger for another two hours of their life.
Does uploading guest audio to a transcription API matter?
More than people assume, and mostly as a promise rather than a regulation. Guests are told 'record locally, send me the file' and they hear that as an assurance. If step one of your pipeline posts that file to a third party, say so up front and check what the provider retains. A local Whisper build removes the question entirely and is fast enough on any recent laptop.
Every week, someone ships something they shouldn’t have.
New verdicts, the worst thing that landed in the trap, and the occasional incident report. No other email, ever.
Cleaning audio is scriptable. Publishing a feed on a schedule is where it becomes a commitment.
Loudness normalisation is a well-specified standard and ffmpeg already implements it.
Transcript to clips and quotes is a prompt chain over Whisper output. Very reachable.
last reviewed 2026-08-05 · verdict is editorial and unsponsored · shared entry data from canivibecodeit under MIT · not legal advice