Should I vibe code
Record or import spoken audio, clean it, transcribe it, and assemble a simple episode
Every episode you have shipped is thirty minutes of clean, isolated reference audio for cloning your guest's voice.
?
Their verdict, the Essentials (the tier formerly named Storyteller) 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
Most of this is the same pipeline the Cleanvoice and Auphonic entries already bless: record, denoise, normalise, transcribe, export. What sets Podcastle apart is the voice model, and it changes the shape of the whole thing. Cloning a usable voice now takes a few seconds of clean reference audio, which means every episode you have ever published is a training corpus for the named people who appeared on it, and the moment you build this yourself, your project directory contains reference clips and embeddings for all of them. The feature that makes it worth building — patching a mispronounced word inside a real sentence, in the speaker's voice — is precisely the feature that produces a recording of someone saying something they did not say, with no disclosure attached and no way to prove afterwards which words were theirs. Then it ships into an RSS feed, which means a thousand podcast apps have cached it inside an hour and there is no recall. Nothing here is beyond a competent weekend, which is exactly the problem. The dangerous half is a pip install; the tedious half — remote multitrack that survives a guest's laptop going to sleep — is what will actually stop you.
What actually breaks
not "if". the specific failures.
- The consent question, which is not one question: consent to be recorded, consent to be published, and consent for a voice to be modelled are three different permissions and most implementations collect one checkbox
- The reference clips and voice embeddings sitting in your storage, which are the highest-value thing in the project and are usually the least protected
- A word patched into a real sentence, which produces audio that is genuinely the speaker's voice saying something the speaker never said, with nothing in the file marking which part is synthetic
- Recall, because an MP3 in an RSS feed is in a thousand podcast clients' caches within the hour and there is no version of deleting it
- Remote recording, the first time a guest's laptop sleeps, a tab refreshes, or the browser reclaims memory mid-session and the local track is gone
- Track alignment, when three guests' local files start at different offsets and every edit computed on one lands mid-word on another
- The master, if any render writes over it — a two-hour interview with someone who will not sit down with you twice has exactly one original
- Reproducibility, since the model that made a clip was pulled from a public host and has been updated twice since, so you cannot show what the file used to be
- Retention, because guest audio is other people's voices and 'we keep everything forever' is a decision even when nobody made it
The guest emailed asking for a correction — he had said twelve percent and meant twenty, and could you fix it. You did, in about ninety seconds, using the text editor: change the word, regenerate it in his voice, splice, render. It sounded perfect because it was his voice. Six weeks later the clip is circulating with that number highlighted, and someone in his employer's compliance function wants to know when exactly he said it and on what basis. The waveform is him. The words are yours. The email where he asked is in a thread you cleared out during an inbox purge. And the model that made the splice was a checkpoint you pulled from a public host in March, updated twice since, so you cannot even regenerate the file to demonstrate what changed. There is no bug here. Everything worked exactly as you built it, including the part where nothing in the exported file records that four hundred milliseconds of it were synthesised.
Is that you?
the verdict is a default, not a law
- It is your own voice, your own recordings, and no synthesis of anyone else
- Consent to record, to publish and to model a voice are three separate stored permissions with dates on them
- Every synthesised segment is watermarked, logged in the edit list, and disclosed in the published episode
- Guests record locally at full quality and the master file is never written over
- You will clone a voice from audio you did not get explicit, specific permission to model
- Synthesised words go into real sentences with no marker in the file and no note in the show
- Guests are in jurisdictions where recording consent rules differ and your app assumes one answer
- Reference clips and embeddings live in the same bucket as everything else, with the same access rules
- There is no story for deleting a guest — recordings, transcripts, reference clips, embeddings — when they ask
If you build it anyway
the checklist, then the prompt that enforces it
- Collect three permissions, not one: record, publish, and model this voice. Store them as append-only records with the date, the wording shown, and who agreed. The third one is the one that matters and the one everyone folds into the first.
- Make the consent record a hard dependency of synthesis in the data model. A voice with no modelling permission should be impossible to clone, not merely discouraged by a greyed-out button.
- Announce recording to every participant at the start of every session and log the acknowledgement. Recording-consent law varies by US state and by country, and your app should not be picking one on their behalf.
- Watermark synthesised audio and write it into the exported file's metadata, along with the model name and version. Future-you needs to be able to explain a clip you no longer remember making.
- Record each participant locally at full quality with resumable chunked upload. The live stream is the fallback, never the master, and recovery of a half-finished local recording is a feature you build on purpose.
- Never render over a master. Edits are a list of operations against an untouched original, so a bad denoise pass costs a re-render rather than an interview.
- Store reference clips and voice embeddings separately, encrypted, with their own access rules and their own retention clock. They are the crown jewels of this system.
- Build guest deletion before you build the second feature: recordings, transcripts, reference audio and embeddings, all of it, on request, and test it.
- Disclose synthetic speech in the episode itself when it is more than a repair. The EU AI Act's transparency obligations are the floor, not the ceiling.
I am building a podcast studio: browser recording with remote guests, cleanup,
transcription, and synthetic voice. The synthesis is the dangerous part and also
the easiest part, so it comes last. Refuse to reorder this.
1. Consent before capture. Before a microphone opens, record who the person is,
what they agreed to, and specifically whether their voice may be modelled.
Three permissions — record, publish, model — stored append-only.
2. Announce recording in the interface to every participant, every session, and
log the acknowledgement. Recording-consent rules differ by US state and by
country; do not hard-code one answer.
3. Record each participant locally at full quality with resumable chunked
upload. The live stream is a fallback, never the master.
4. Treat a sleeping laptop, a tab refresh and a reclaimed tab as expected
events, and make recovery of a partial local recording a first-class path.
5. Never render over a master. Every edit is an operation list against an
untouched original file.
6. Only now, synthesis. A voice may be cloned only where the consent record
permits modelling. Enforce that in the data model, not in the interface.
7. Refuse to clone a voice from a file I upload with no matching consent
record, and tell me why instead of proceeding quietly.
8. If I ask you to patch a word inside a real sentence, first tell me plainly
that the output is a sentence the speaker never said, then require an
explicit approval that you log.
9. Watermark every synthesised segment, write the model name and version into
the exported file's metadata, and keep which words were generated in the
edit list.
10. Build guest deletion early: recordings, transcripts, reference clips and
embeddings. Someone who withdraws must leave nothing behind that can still
speak in their voice.
11. Out of scope unless I ask: publishing to an RSS feed, dubbing, lipsync.
Finish by telling me what this costs to buy instead.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
The moment another human being's voice is involved. Twenty dollars a month buys a vendor who has written terms about voice cloning, has a consent flow, watermarks output, and can be pointed at when a guest asks what you did with their recording. That last part is not software and you cannot write it. Build the recorder and the cleanup chain for your own voice by all means — that is a good weekend and the Auphonic entry explains why. Buy the part that synthesises other people.
$19.99/mo is cheaper than your weekend.
Three things have to survive the application, and only one of them is audio. Keep the master recordings as plain files in dated folders — one per participant per session — because those are irreplaceable and everything else is derived. Keep the transcripts with timestamps as text next to them. And keep the consent records somewhere you keep contracts, not somewhere you keep databases, because the question they answer arrives years later and arrives in writing. On the way out, the last task is not archiving the repo: it is deleting reference clips and voice embeddings for everyone who is not you, and being able to say that you did.
Local speech recognition with word-level timing, which is the transcription and text-editing half done.
Open voice cloning from a short reference clip — included here because it demonstrates how little effort the risky feature now takes.
Long-running open-source multitrack audio editor and useful implementation prior art for the editing half.
Questions
Is Podcastle still called Podcastle?
No. It rebranded as Async in February 2026 and podcastle.ai now redirects there; the entry keeps the old name because that is the name on the sibling site's listing and the one people search for. The product line is unchanged and the tiers kept their structure — the payments endpoint still calls the Essentials SKU 'monthly_storyteller' — but the positioning has widened from podcasting to a general AI video, audio and voice platform. If you are pricing it today, you are pricing Async.
Cleanvoice and Auphonic are DEMO ONLY and SHIP IT. Why is this a band worse than both?
Because they process a recording and this one manufactures one. Auphonic's entry is loudness and levelling — arithmetic on a file you own. Cleanvoice's is about cuts that measure right and sound wrong, which is an aesthetic failure you carry alone. Podcastle adds a model that produces new speech in a real, named person's voice. That moves the harm off your show and onto someone who agreed to be interviewed, and it does it in a medium where the output is indistinguishable from the original.
What is actually illegal about cloning a voice?
It depends where everyone was sitting, which is the problem. In the US, right-of-publicity statutes cover a person's voice and several states have recently extended them explicitly to digital replicas — California's AB 1836 among them. In the EU, the AI Act imposes transparency obligations on synthetic audio, and the underlying recording is personal data needing a lawful basis in the first place. None of that turns on whether your intent was good. It turns on whether you can show what you were permitted to do, which is a records problem you either designed for or did not.
Can I at least clone my own voice?
Yes, and that is genuinely the good version of this build. Your own voice, your own recordings, your own machine, no publishing pipeline: the failure modes shrink to a bad render and a wasted afternoon. Keep it there and the entry's verdict does not really apply to you — it applies the moment a second person's voice enters the project.
- RCFP reporter's recording guide — US state-by-state consent (US)
- GDPR Art. 6 — lawfulness of processing (EU)
- EU AI Act Art. 50 — transparency obligations for synthetic audio and deepfakes (EU)
- California AB 1836 — use of digital replicas (US, California)
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.
Cutting the 'ums' is one afternoon. Making the cut inaudible is the entire product.
Loudness normalisation is a well-specified standard and ffmpeg already implements it.
The show notes are disposable. The timestamped transcript is the archive, and it belongs in your repo.
last reviewed 2026-08-05 · verdict is editorial and unsponsored · shared entry data from canivibecodeit under MIT · not legal advice