Using voice with Claude Code: record a memo instead of dictating

Most of what I hand Claude Code started out as speech, and almost none of it goes through Claude's own dictation. I hit record on the Mac, talk for as long as I want, stop, and a few seconds later the transcript is sitting there for me to paste in. That is the loop, and it is for work I am doing right now with Claude Code open in front of me.

I did expect to be using /voice for this and I am not, which took me a while to admit.

Why the built-in dictation didn't stick

Claude Code's /voice is real dictation and it works: hold space, talk, and your words appear in the prompt as you speak. I still use it, but only for the shortest things, a line or two where I would otherwise be typing a sentence. Past that it stops being worth it. It takes longer than you expect, it mishears enough that I am editing the result instead of reading it, and I cannot predict which of those two I am going to get.

Some of that is visible in how it is built. Your audio goes to Anthropic's servers rather than being handled on your machine, and in tap mode the recording stops on its own after two minutes or after fifteen seconds of silence. Two minutes is a reasonable ceiling for a prompt and nowhere near enough for thinking something through.

Then there is language, and this one took me a while to work out because it looks like the dictation simply can't do Swedish. It can. Swedish is one of the twenty it supports. What it doesn't do is detect anything: dictation uses whatever language you have configured, and if you haven't configured one it assumes English and transcribes your Swedish as though it were English, which produces the sort of output that makes you conclude the feature is broken. Anthropic's own troubleshooting page lists this under transcription coming back garbled or in the wrong language.

You can set it, and here's the catch: it's the same setting that controls what language Claude replies in. So dictating in Swedish means being answered in Swedish, and if you go back and forth between languages during a day, as I do, you are changing a config setting each time rather than just talking. In practice everybody leaves it on English and puts up with the results.

Whisper detects the language itself, per recording, with nothing to configure and no effect on anything else. Finnish comes back Finnish, and the memos that wander between two languages halfway through, which most of mine do, come back intact. Finnish also isn't among the twenty at all, which where I live settles it.

The loop I actually use

I start a voice memo on the Mac, right there at the desk, in the middle of working. Then I talk for as long as it takes, which is usually a few minutes and sometimes twenty, while I page through the code or stop to read a diff or lose my thread completely and start the sentence again. Nothing is counting down and nothing needs to stay in the foreground. When I stop, My Transcriber transcribes it on the Mac and the text is there almost immediately, and I paste it into Claude Code.

"Almost immediately" is the part that makes this work rather than a nice idea. Whisper on an Apple Silicon Mac runs at about thirty times faster than real time with Metal, so ten minutes of talking comes back in around eighteen seconds and an hour takes under two minutes. Wait long enough to reach for your coffee and it's done.

The other one, which is genuinely away from the desk

Separately from all that, I record memos on my Apple Watch or my phone — my own thoughts, a conversation, a meeting I'm sitting in. Those sync over iCloud, get transcribed on the Mac whenever they land, and are waiting as markdown by the time I sit down. It's the same mechanism and a different habit, and it isn't the hands-on Claude Code loop above. It's how the thinking I did on a walk ends up somewhere I can use it.

Then let Claude Code go and find them

Pasting is fine and I do it constantly, but you can skip it. My Transcriber ships an MCP server, so Claude Code can query the voice memo index directly. I say "I talked about this in one of my recent voice memos, go and read it," and it does. The app sets that up on first launch by writing an .mcp.json into your data folder along with a bundled skill, so running claude from that folder picks it up.

It works the other way round too. Claude Code can put things into the library rather than only reading from it, so "download the recording of yesterday's meeting and transcribe it" runs Whisper on your Mac and the recording turns up alongside your voice memos, searchable, in the app. Ask Claude Code to transcribe something on a machine without My Transcriber and it will still have a go — usually by installing Whisper itself and improvising a script around it, which does work and leaves you with a stray text file that nothing else knows about.

Four things that make this work better

Don't bother introducing it. I went through my own library while writing this and almost none of my memos say what they're about — they open mid-thought, halfway into a sentence, the way you actually start talking when something occurs to you. It doesn't matter, because the whole transcript is searchable and describing the content is enough to find it later. A spoken subject line would be a chore you'd give up on by the third memo.

Don't perform, either. It's going to a language model, not to a reader, so false starts and repetition and changing your mind halfway through a sentence are all fine. The messy version is often the useful one, because the reasoning is in there rather than just the conclusion.

Type the short ones. The split that matters isn't voice against typing, it's short against long. A one-line instruction is quicker typed than spoken, and that's the job dictation is for if you want it. Everything past a couple of sentences is faster recorded, and gets more accurate rather than less as it goes on.

Let them pile up. A single transcript is a note. A year of them is context you can ask questions across, which is a different thing entirely, and it costs nothing to keep because they're small text files sitting in a folder you already back up.

Setting it up

1. Install My Transcriber and grant Full Disk Access, which is what lets it read Voice Memos synced from iCloud.
2. Next time you're about to type a long thing at Claude Code, record a voice memo on the Mac instead and talk it through. Stop when you're finished; the transcript appears in your folder within seconds.
3. Paste it in. That's the loop, and it's most of what this is for.
4. Later, once memos have accumulated: run claude from that folder and ask it about them directly, and record on your phone or Watch when you're away from the desk.

Common questions

Related: using voice memos as context for Claude, ChatGPT and Obsidian, and the full documentation.