Writing about speech, text and privacy on a Mac
What the built-in tools actually do, where they stop, and how to tell a claim about local processing from a claim about local recording.
Free transcription on a Mac: what each free route costs instead
Four ways to transcribe on a Mac without paying, measured on the same recordings: what each costs in disk, minutes and accuracy. The smallest model is 141 MB.
MacWhisper alternatives, and when it is already the right answer
MacWhisper Pro is €64 once and the free tier is not a demo. What the paid version adds, where the alternatives really differ, and why none transcribe better.
Mono, stereo, or two files: what actually survives when people talk over each other
Putting two speakers in separate stereo channels does nothing for transcription — I measured identical output. When they overlap, a whole reply disappears.
superwhisper: what the free tier covers, and what $8.49 adds
superwhisper's free tier includes meeting recording and unlimited local Whisper. File transcription is the paid line. Three price shapes, and which one wins.
Where the delay goes between the key press and the pasted text
A one-second phrase costs Whisper 711 ms and a 26-second one costs 866 ms. The encoder is the same size either way, and that explains most of dictation latency.
How to measure word error rate on your own audio
Published accuracy figures are measured on somebody else's recordings. Here is the script, the reference problem, and why the same transcript scored 23% and 29%.
QuickTime screen recording with audio: what the Mac will and will not give you
QuickTime records your screen and your microphone, never the sound the Mac is playing. Here is the proof from the system itself, and the three real workarounds.
What the licence of a local speech model actually lets you do
Three models, three licences: MIT, CC BY 4.0 and one you have to read. None says a word about the transcript. What they cover, and where the answer lies.
Two engines, one recording, and 130 ms of disagreement about when a word starts
Whisper and Parakeet transcribed the same file. On 41 matched words their timings differ by 130 ms at the median and 460 ms at the worst. Why, and when it matters.
Why Whisper says "Thank you" to an empty room
Four Whisper models transcribing pure digital silence. Two of them thank you for it, one loops a sentence eight times. The runs, and what actually stops it.
I gave an agent a folder of recordings and let it fix its own transcripts
Three standup recordings, twenty-five word errors, and a custom vocabulary the agent wrote itself from what it got wrong. Error rate went from 12.5% to 4.0%.
AI video dubbing is four steps, and the first one decides the rest
Dubbing a video is transcription, translation, voice synthesis and timing. What breaks at each step, why translated text grows by 130%, and what to check.
I compressed a recording 22 times to see how much accuracy it costs
A 2 MB recording squeezed to 97 KB cost three percentage points of word error rate. The measurements, the method, and the one result that surprised me.
I built an MCP server with twelve tools. Here is what the tutorials leave out.
Six lessons from shipping a local MCP server: why the tool list guards nothing, why a matching user ID proves nothing, and how clients truncate you silently.
Call recording laws: what the statutes actually say, quoted
Six statutes quoted verbatim, from the US federal rule to California, Washington and Canada. Why published lists of all-party states disagree with each other.
Extract audio from a video on a Mac without re-encoding it
Two of the three ways are already on your Mac, and one of them copies the audio track bit for bit instead of re-encoding it. Commands and measurements.
A local model against a per-minute API: where the line actually falls
A speech to text API costs $0.003–0.006 a minute. A local model costs a download. The arithmetic, including the range where the API is the better buy.
Your Mac already has 184 voices. Here is how to use them
macOS ships a text to speech engine with 184 voices in 51 languages, works offline and writes audio files from the command line. Where it stops.
Voice coding changed meaning, and the new one actually works
Speaking code aloud never worked well — code is not natural language. Speaking to an agent is a different job, and it is the one dictation was built for.
AI note takers: the bot is not the privacy question
No bot and no AI training are separate promises, and the vendors that make one often break the other. What Otter, Fathom, Granola, Fireflies and Zoom say.
Dragon for Mac is gone. What actually replaced it
Nuance discontinued Dragon for Mac on 22 October 2018 and stopped updating it. What people actually used it for, and which of those jobs modern Mac apps do.
The Mac dictation shortcut, and the two reasons it stops working
How to set the dictation shortcut on a Mac, and what to do when it silently does nothing — Secure Input and the offline language table macOS keeps on your disk.
"On-device" transcription: how to check the claim, and what it costs
Offline speech to text is a checkable claim, not a promise. Three ways to test any app in a minute, and the disk and startup cost nobody advertises.
Otter.ai alternatives, grouped by why you are leaving
Otter caps the free tier at 300 minutes and three file imports ever. Its competitors, sorted by the reason you are leaving: limits, price, the bot, or training.
Transcribe a voice memo on a Mac: the free way, and where it stops
Voice Memos transcribes for free on Apple silicon, in ten languages, for recordings made inside the app. What that covers, and what to do when it doesn't.
Transcribing interviews for qualitative research, end to end
Verbatim or clean verbatim, what ethics approval says about upload, and which of NVivo, ATLAS.ti and MAXQDA takes your file without converting it first.
Which model transcribes your language, and which vendor is rounding up
Transcribe Spanish audio, or Japanese, Korean or Russian: three local models cover different sets, and one vendor's count does not survive a look inside.
"Translate audio to text" means two different things
Most people searching this want a transcript, not a translation. How to tell which you need, and why the fastest Whisper model ignores the translate flag.
Whisper, Parakeet or SenseVoice: numbers, not adjectives
Whisper transcription is the default local answer, and often the wrong one. Word error rates and a same-machine speed run for five models on one Mac.
Wispr Flow alternatives for Mac
What Wispr Flow costs over three years, why its transcription is always in the cloud, and the Mac alternatives that charge once. With superwhisper compared.
How to record a FaceTime call
macOS records a FaceTime audio call for you, with a notice both people hear. Where that stops, why screen recording captures no sound, and what to do instead.
Best dictation apps for Mac in 2026
Eight ways to dictate on a Mac, compared on the two things that matter: where your speech is recognised, and whether you pay once or every month.
How to transcribe audio and video to text on a Mac
Four ways to transcribe audio and video to text on a Mac: Voice Memos, online services, local transcription software, Whisper. Costs, formats, privacy.
Voice to text on a Mac: four ways, compared
Four ways to do speech to text on a Mac: built-in Dictation, Voice Control, paid apps and Whisper. What each does, where each breaks, how to pick.
How to get a transcript from any YouTube video
Transcribe from YouTube for free using YouTube's own transcript panel. Where auto captions fail, what the terms allow, and when to convert video to text locally instead.
