Should I vibe code
Transcription, translation, subtitles, and collaborative media review
Transcribe, translate, burn in, publish. Four steps, and after step one nobody who can hear the audio is looking.
?
Their verdict, the Advanced 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
Give Sonix credit for what separates it from every notetaker on this site: your software never opens a microphone. The file turns up already recorded, and whoever recorded it made the consent decision before your code existed. That retires the argument seven other entries are having, and it is why the safe version of this build is unusually good — whisper.cpp on your own machine, files that never leave it, and you have a transcription service that beats the paid one on the axis that actually matters. The regret is further down the pipeline. Sonix does not stop at a transcript. It translates into fifty-five languages, generates SRT and burned-in subtitles, and hands you a share link, and every step past transcription removes one more person who might have noticed that the model filled four seconds of room noise with a fluent, confident sentence nobody said. By the time that sentence is a Spanish caption baked into a published video it is a quotation attributed to a named human being, distributed, and unfixable. Put a media archive nobody ever deletes underneath it and the numbers land where they land.
What actually breaks
not "if". the specific failures.
- The model on silence and background noise, which emits fluent invented sentences rather than an inaudible marker — and does it most in the low-quality stretches you most wanted help with
- Speaker labels, which are a guess. Two similar voices merge, and one person's sentence is attributed to another in the document everyone quotes from
- Whatever hosted API you piped the audio through, which has just re-disclosed somebody else's interview to a company they were never told about
- Timing drift on long files, so subtitles are correct at minute two and two seconds late by minute forty, which is where they stop being subtitles and start being distracting
- Translation, where the error goes invisible: nobody on your team reads the target language, so nobody can notice the caption now says the opposite
- Burned-in subtitles, which are pixels. Correcting one means re-encoding and re-uploading, and every copy already downloaded keeps the original mistake
- The archive. Media files are large, deletion is nobody's job, and an ad-hoc tool quietly becomes a searchable corpus of named people's recorded voices with no retention policy and no owner
The customer testimonial was recorded in a hotel lobby, which is why eleven seconds of it are mostly air conditioning. The transcript reads straight through them, because the model has no way to shrug: it filled the gap with a sentence about switching away from a competitor — plausible, in her register, entirely invented. Nobody re-listens, because the transcript is the thing people re-read. It goes out for translation, comes back as Spanish subtitles, gets burned into the cut and published on Tuesday. Ten days later her legal team asks where the claim came from. There is no correction to publish, because the words are pixels in a video that has already been reposted twice, and the one person who could have caught it — somebody who both heard the room and reads Spanish — never existed anywhere in your pipeline.
Is that you?
the verdict is a default, not a law
- The audio is yours, or recordings you already hold lawfully, and the transcript is for your own reading
- It runs locally and no media file leaves the machine
- The output is treated as a draft that someone checks against the audio before it is used for anything
- You wrote down a retention period and the delete button genuinely deletes
- The recordings are clinical, legal, journalistic or research interviews
- Transcripts get quoted, captioned or published without anyone re-listening to the relevant passage
- Anything is translated into a language nobody on the team reads
- Share links go to people outside the team, or never expire
- Audio is uploaded to a hosted API the speaker was never told about
- The media archive has no retention policy, which after six months means it has one: forever
If you build it anyway
the checklist, then the prompt that enforces it
- Run the model locally. whisper.cpp or faster-whisper on your own hardware removes the entire third-party disclosure problem, and for file-based transcription the quality gap is small enough not to be the deciding factor.
- Surface confidence. Mark low-confidence and no-speech segments visibly in the output instead of emitting clean prose over a silent stretch — the fluent sentence over silence is this tool's signature failure.
- Require a human pass against the audio before a transcript is quoted, captioned or sent anywhere. Make that a workflow state, not a habit.
- Never translate without a reviewer who reads the target language. An unreviewed translation is an unverifiable claim in somebody else's mouth.
- Prefer sidecar SRT or VTT over burned-in subtitles. Burn-in is the only genuinely irreversible output in this pipeline.
- Set a default retention on media and transcripts, delete on schedule, and check that delete removes the object rather than hiding a row.
- Treat share links as personal data access: expiring, revocable, logged, and scoped to one file.
- Strip or redact obvious identifiers before anything leaves the workspace, and keep the unredacted copy in one place you can point at.
I am building a file-upload transcription tool: audio and video in, transcripts,
subtitles and translations out. The recordings are of real people who are not in
the room. Apply these before any UI work and push back when I ask to skip one.
1. Transcription runs locally by default — whisper.cpp or faster-whisper. Do
not add a hosted speech API unless I ask twice, and when I do, tell me
plainly that the audio is being disclosed to a third party the speaker
never heard of.
2. Preserve and expose per-segment confidence and no-speech probability. Any
segment above the no-speech threshold is rendered as an explicit inaudible
marker, never as text. Silence must never become a sentence.
3. Build the review step before the export step. A transcript starts in an
unverified state, and only a human who listened can move it to verified.
Exports refuse to run on unverified content by default.
4. Speaker labels are marked as inferred everywhere they appear, and I can
correct them. Never print a name next to a sentence as though it were fact.
5. Retention first: a configurable default lifetime for media and transcripts,
a scheduled purge, and a delete path that removes the file from disk and
from any derived artefacts. Prove it with a test.
6. Subtitle output defaults to sidecar SRT/VTT. If I ask for burn-in, warn me
that it cannot be corrected after publication and make me confirm.
7. Translation requires an explicit reviewer field naming who checked it. If
nobody reads the target language, refuse and say so.
8. Share links expire, are revocable, are logged, and are scoped to one file.
No directory listing, no guessable ids.
9. Never log transcript text or file names into application logs, and keep
uploads off any path served statically.
10. Out of scope, and say so rather than half-building: real-time
transcription, diarisation you would trust, and anything resembling a
compliance claim. If the recordings are clinical or legal, buy a service
that will sign paperwork about them — Sonix is $10 an hour.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
When the recordings are somebody else's and the output is going to be published. Pay-as-you-go at $10 an hour means casual use costs less than an evening of your time, and what the subscription really buys at $50 a month is the editor, the translation pipeline and a vendor who will sign paperwork about how they handle the audio. If your files are clinical, legal or research material, that paperwork is the product and you cannot generate it yourself. Build your own when the audio is yours, the transcript is for you, and it never leaves your laptop — that version is genuinely better than paying.
$50/mo is cheaper than your weekend.
The transcripts are the easy half: they are text, SRT and VTT, and every tool in the world reads them. The media is the part with a tail. Decide now how long you keep source files, write it down, and enforce it, because the default outcome of a homemade transcription tool is a disk full of other people's voices that outlives the reason it was collected. When you shut the project down, exporting the transcripts is an afternoon and deleting the audio is the actual work — including the copies in your backups, which is the bit everyone forgets.
Local speech-to-text that runs fast enough on a laptop to make the private version of this tool the easy one.
Adds word-level alignment and speaker diarisation, which is exactly the part naive Whisper output gets wrong on long files.
Self-hosted transcription and subtitle-editing suite shaped much like Sonix; the README says the current branch gets no further releases pending a rewrite.
Questions
If the user supplies the file, whose consent problem is this?
Not yours at capture, which is the honest and unusual thing about this entry — no bot joins a call, no microphone opens. It becomes yours afterwards. You are the one storing recordings of identifiable people, deciding how long to keep them, deciding who a share link reaches, and producing a written record of what they said that other people will rely on. Consent covers the recording. It does not cover a sentence they never spoke appearing under their name.
Why is this harsher than the Notta entry?
Notta is rated as a personal archive: your recordings, your transcripts, your problem. Sonix's pipeline keeps going after the transcript — translation, captions, burn-in, shared links, a team — and every one of those steps takes a human further away from the audio while making the output harder to retract. Same model, same hallucination, considerably more distance between the error and anyone able to spot it.
Is a local Whisper build actually good enough?
For file-based work, usually yes, and it is better on the dimension this entry cares about: the audio never leaves your machine. What you do not get for free is the tedious surrounding product — the synchronised editor, clean speaker labels, translation, subtitle timing, export formats that video editors accept. That is a lot of afternoons, and it is what the $10 an hour is actually buying.
- GDPR Art. 5 — principles, including accuracy of personal data (EU)
- GDPR Art. 32 — security of processing (EU)
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.
Whisper never says "I couldn't hear that". It writes a plausible sentence instead, and your archive keeps it forever.
Text-based editing is a beautiful idea sitting on top of very fiddly media plumbing.
A summariser wired into your ATS turns one misheard sentence into a permanent note on a candidate's file.
last reviewed 2026-08-05 · verdict is editorial and unsponsored · shared entry data from canivibecodeit under MIT · not legal advice