How to transcribe video on a Mac, and what to do with the transcript
Getting accurate transcripts offline, choosing between TXT, SRT and DOCX, and the second life a transcript has once you stop thinking of it as subtitles.
Transcription used to be either expensive or bad. You paid roughly a dollar a minute for a human, or you got automatic output that took longer to fix than to retype.
That is over. On an Apple Silicon Mac, speech to text now runs on-device, offline, at no per-minute cost, with word-level timecodes, in dozens of languages, at an accuracy that beats a fast typist on clean audio. It is the single most mature piece of the whole AI-and-video story.
What most people still miss is the second half: what a transcript is good for once you stop thinking of it as subtitles.
What runs locally, and how well
The models behind this are in the Whisper family: open speech recognition models that come in several sizes. On Apple Silicon they run on the GPU and Neural Engine, which is why a laptop handles them comfortably.
Practical expectations:
- Clean interview audio, one speaker, decent mic: very high accuracy. Errors cluster in proper nouns and technical terms.
- Two people, occasional overlap: good, with the overlap moments getting messy.
- Noisy location audio, wind, café background: degrades noticeably, still usually better than nothing.
- Heavy accents or code-switching between languages: variable. Test on your own material before promising a client anything.
- Speed: faster than real time on modern hardware, often several times faster. A 40 minute interview is a coffee break, not an afternoon.
The trade between model sizes is the usual one. A small model is quick and slightly sloppier. A larger one is slower and better on hard audio. For search purposes, the small one is usually fine, because “did anyone say revenue” does not need perfect punctuation. For a client deliverable, use the larger one.
One systematic quirk worth knowing: speech models are trained to produce clean, readable text, which means they quietly drop filler sounds. Your “um” and “äh” often do not appear in the transcript at all, even though they are audibly there. That is helpful for a readable document, and it is a problem if you are trying to use the transcript to locate hesitations.
Choosing an export format
Get more than one. They are cheap and they serve different jobs.
| Format | Use it for |
|---|---|
.txt | Search. Spotlight indexes it, so your spoken words become findable from the desktop. |
.srt | Subtitles, and importing timecoded text into an NLE. |
.docx | Sending to a human. Review, comments, client sign-off. |
.pdf | Handover and archive, when the layout should not move. |
The one people underestimate is .txt. Saved next to the clip, it turns your spoken content into part of the same Spotlight index as everything else on the Mac. A search for a phrase you half remember from an interview eighteen months ago returns the transcript file, sitting in the folder next to the video. That is moment-level search of your archive, for the cost of saving a text file. More on that pattern in finding clips without Premiere.
Naming and filing transcripts
The rule: a transcript lives next to its clip and shares its name.
Rooftop-Interview-Marcus_16x9_4K_25FPS.mp4
Rooftop-Interview-Marcus_16x9_4K_25FPS.txt
Rooftop-Interview-Marcus_16x9_4K_25FPS.srtNot in a Transcripts folder three levels away, and not with a name that only matches by accident. When a search hits the text file, you want the video to be one glance away, and when you archive the project, you want them to travel together.
If you keep a separate transcripts folder for client delivery, make it a copy. The working set stays next to the media.
The uses beyond subtitles
This is where the value actually is, and it is the part most workflows never reach.
Finding the moment. Timecoded transcripts mean “where did she talk about the supply chain” is a text search that returns a timecode. For interview-heavy work this is the single biggest time saving available anywhere in post.
Naming clips. The first sentence of a talking clip is usually an excellent description of it. A clip that starts “so we founded the company in 2019” is obviously the origin story clip, and it can be named that automatically.
Working out the structure before you edit. With every interview transcribed, you have the entire content of the shoot as text, on one screen. Reading 12,000 words takes twenty minutes. Watching the same material takes four hours. You can see the shape of the piece, spot the repeated answer, notice that the best line came at the end of a rambling take, and plan the story before you open a timeline.
Feeding it to something else. A transcript is portable text, so it can go into whatever you use to think: a document, an outline, a summarizer, a search over a whole archive. The transcript is the raw material for structure, and structure is where the actual work of a piece is.
Repurposing. Blog version, show notes, chapter markers, quote cards, captions. All of them start from text you already have.
Accessibility and reach. Captions are not optional for a lot of publishing, and platforms index caption text, which affects whether anyone finds the video at all.
The one thing to fix by hand
Proper nouns. Every time.
Names of people, companies, products and places are where every speech model, cloud or local, makes its errors. It is a two minute pass: search and replace the four names that recur, and the transcript goes from useful to publishable.
If you transcribe the same subject repeatedly, keep a small list of the terms that get mangled and how they should read. It stays useful across projects.
Where this fits in the ingest pass
Transcription belongs at ingest, not at edit time. The reason is the same as with tagging: at ingest you are getting to know the footage anyway, and after the edit you will never come back.
An order that works:
- Copy the card, verify.
- Analyze: descriptions and subject tags for everything, transcripts for anything with speech.
- Rename and file, using both.
- Read the transcripts. Plan the piece.
- Open the NLE, knowing what you have.
Step four is the one that feels like a luxury and turns out to be the time saver. Editing is much faster when you already know what is in the material.
Audio recordings, not just video
If you record separate audio, a lav recorder, a room mic, an interview on a handheld, the same applies and the payoff is bigger, because an audio file has no thumbnail at all. ZOOM0007.WAV is even more opaque than C0087.MP4.
Transcribing and describing recordings turns a folder of numbered files into a library. Organizing audio recordings covers that case, including sound effects and ambience where there is no speech to transcribe.
Cliptag transcribes on-device on macOS as part of the ingest pass: drop in a folder, and every talking clip and every recording comes back with a transcript, exportable as TXT, DOCX, PDF or SRT with timecodes. The transcript is saved next to the clip so Spotlight indexes it. On-device mode on Apple Silicon is free and unlimited.
