Last month I ran the same 58 minute episode through three different AI transcription tools and then sat down and typed out the first twelve minutes by hand. The machines agreed with each other most of the time. All three of them were wrong in the same place, and it was the place that mattered.
The host asked a question. The guest paused for about four seconds, laughed once, and said "yeah, no." Every tool turned that into "Yeah. No." with a full stop in the middle, which reads like agreement followed by contradiction. What actually happened in the room was a person being asked something uncomfortable and dodging it with a laugh. The words were technically captured. The meaning was gone.
That is the whole argument I want to make here. AI transcription is very good at words and completely blind to intent. If you build your edit on top of the transcript, you inherit the blindness.
The accuracy number is a lie of omission
Vendors quote word error rates. Ninety five percent, ninety seven percent, sometimes higher on clean audio. Those numbers are real and I do not think anyone is cooking the books. The problem is that word error rate treats every word as worth the same amount.
Take a 60 minute episode. Roughly 9,000 words spoken. At a very respectable 3 percent error rate that is 270 wrong words. Now ask yourself which 270. In my experience they cluster in exactly the worst places:
- Proper nouns. Company names, product names, the guest's own surname.
- Numbers with units. "fifteen K" becomes "fifty K" and nobody catches it until a viewer does.
- Crosstalk. Two people talking over each other, which is usually the most alive moment in the episode.
- Accented speech, especially non-native English, which is most of the interviews I work on out of Dubai.
- Anything said quietly, which is often the honest part.
So the 97 percent is spread across "and", "the", "you know", and the 3 percent lands on the guest's company name and the one statistic they came on the show to say. That is not a rounding error. That is the episode.
Timestamps drift, and the drift is invisible
The second problem is worse because it does not look like a problem. Word level timestamps from AI models are estimates. On clean single speaker audio they are usually within about 40 milliseconds, which is fine. On overlapping speech or after a long stretch of music they can slide by a quarter second or more.
A quarter second does not sound like much until you cut on it. Cut a sentence using a drifted timestamp and you clip the first consonant off the next word, or you leave a stray breath hanging at the head of the clip. The transcript says the cut is clean. Your ears say something is off and you cannot tell what. I have watched editors spend twenty minutes hunting a problem that came entirely from trusting a number instead of a waveform.
This is why we never let a tool cut by silence detection. A human editor watches the whole episode, start to finish, and every cut is a decision somebody made on purpose. It is slower. It is also the reason the revision round on our shows tends to be about taste rather than about fixing damage.
What the transcript cannot see
Here is a partial list of things that never make it into a text file:
A guest looking away before answering. The half second where the host decides not to interrupt. A hand going up. Somebody leaning back when they get comfortable. The exact frame where a laugh becomes genuine instead of polite. The moment the energy in a remote call drops because one person's connection stuttered and they lost the thread.
None of that is text. All of it is editing information.
A transcript tells you what was said. It does not tell you what was meant, and it never tells you what to keep.
The best moment in most episodes is not a quotable sentence. It is a reaction. If your workflow starts with reading, you will select for quotability and you will systematically throw away reaction. That is one reason so many AI assembled episodes feel technically fine and emotionally flat. It is also why the strongest moment usually shows up around minute 42, well past the point where a keyword search would have told you to look.
Where I do trust the machine
I want to be fair, because I use these tools every single day and I would not give them up.
AI transcription is excellent for:
- Search. Finding the one place a topic came up in a 90 minute recording, in two seconds instead of ten minutes of scrubbing.
- First pass show notes. Not the final copy, but a skeleton I can rewrite. We charge $35 for show notes and chapters, and honestly the machine does maybe forty percent of that job.
- Rough chapter boundaries. Topic shifts are one thing models genuinely detect well.
- Clip candidates. Give me twenty suggestions and I will throw out sixteen. Getting four usable starting points for free is a real win.
- Accessibility captions, after a human corrects the proper nouns.
Notice what all of those have in common. They are drafts and search indexes. None of them are the edit. The moment the transcript starts driving the timeline instead of informing it, quality drops in a way clients feel but cannot name.
There is a version of this same trap all over my editing stack, where every tool is fine on its own and the danger is only in what you let each one decide.
Remote audio makes it measurably worse
If you record over a call platform, the transcript degrades in ways you would not predict. Compression artifacts eat consonants. Latency means both people talk over each other more than they would in a room, and overlap is where models fall apart. Local track recording fixes most of this, which is the entire reason I keep pushing hosts toward isolated tracks per speaker instead of a single mixed file.
We wrote up the whole approach in making a remote recording not sound remote, and the transcription quality improvement is a side effect people do not expect. Separate tracks mean the model transcribes one voice at a time, and suddenly the crosstalk problem mostly disappears. If you are still choosing a platform, our comparison of Zencastr and Riverside goes through what each one does to your audio before you ever get the file.
The workflow I actually run
This is the honest version, not the tidy one.
- Transcript gets generated as soon as the files land. I read it once, fast, purely to build a map. Where does the conversation change direction. Where does the guest get animated.
- I watch the episode. All of it. No scrubbing. This is the part no tool replaces and it is the part clients are actually paying for.
- I mark cuts on the waveform and the picture, not on the text. The transcript is open on a second screen as a lookup table, not as a timeline.
- Anything I pull for a vertical clip gets its in and out points set by ear, then checked against the text.
- Proper nouns get corrected by hand before show notes are written. Every time. It takes six minutes and it is the difference between a professional deliverable and something a guest quietly resents.
First cut goes out in 48 hours. Revisions come back the next working day. That timeline only works because the human pass happens early rather than as a cleanup step after an automated assembly. Fixing a machine edit takes longer than making a real one, which is the thing nobody tells you when they sell you the time savings.
The pitch for AI editing is that it is cheaper. It is. If you want an episode assembled by a transcript for very little money, that option exists and I am not going to pretend otherwise.
What we charge is $199 an episode for a single camera 1080p show with one revision, and $299 for two to four cameras in 4K with unlimited revisions. Up to 60 minutes as standard. Thirty minute shows cost less, ninety minute shows cost more. Book four episodes and it drops 15 percent, go monthly and it drops 25 percent. Vertical clips are $25 each, show notes and chapters $35, and rush 24 hour delivery is $75 when someone has a launch date. All of it is on the pricing page with no discovery call required.
The difference you are buying is not features. It is judgement. A human decided every cut, which means someone is accountable for the moments that got kept and the ones that did not.
What this means for the first two minutes
Openings are where transcript driven editing does the most damage. The model has no idea that the third sentence is the actual hook and the first two are throat clearing. It will faithfully preserve all three because all three are words.
I cut openings backwards. Find the strongest thirty seconds anywhere in the episode, decide whether it can survive without setup, and build forward from there. More on that in what the first 90 seconds of a video podcast has to do. And when the episode is packaged for YouTube, the thumbnail carries as much weight as the cut does, which is why our sister studio ClickTheory does 8 thumbnails for $250 rather than treating them as an afterthought.
Use the tools, do not obey them
I am not a purist. I would be slower and poorer without AI transcription, and I would rather lose a lot of things than lose transcript search.
But a transcript is a map, and maps are not the ground. The second you let one make decisions about what stays in an episode, you have handed the most human part of the job to something that has never heard a room go quiet.
Read the transcript. Then close it and watch the tape.
If you want a second pair of eyes on a show you are already making, or you are just curious what a human first cut sounds like against your current process, mail me at samy@pixelsamy.com. I am in Dubai and work with hosts wherever they are. Happy to just tell you what I would have cut.