When I first started using voice notes, about half my transcripts needed editing. Names were wrong. Technical terms came out garbled. Sentences ran together into walls of text.
I thought the problem was the transcription engine. Turns out, the problem was my technique.
After testing different approaches for a month, I got my error rate down to less than 5%. Here's what actually made the difference.
Start with a title
Your first few words become the filename. If you don't say a title, you get automatic timestamps like 2025-01-15-143022.md. Good luck finding that later.
Instead, say "Title:" followed by 2-4 descriptive words:
- "Title: Meeting notes with Sarah"
- "Title: Bug report - login flow"
- "Title: Weekend reflection"
This one habit transformed my notes from unsearchable to instantly findable.
Use short sentences with natural pauses
I used to speak in long, winding sentences with multiple clauses that ran together because I was trying to capture everything in one breath without stopping.
That sentence you just read? Terrible for transcription.
The transcription engine uses your pauses to figure out where sentences begin and end. When you ramble without pausing, it guesses. And it guesses wrong.
Now I speak like I write for the web: Short sentences. Natural pauses. Clear breaks.
The transcripts come out formatted correctly without me having to add punctuation manually.
Don't over-specify punctuation
I wasted two weeks saying "comma" and "period" out loud. My transcripts were littered with the literal words "comma" and "period."
The engine adds punctuation based on your cadence and phrasing. When you pause briefly, it knows to add a period. When your voice goes up slightly, it knows it's a question.
Just speak naturally. Let the engine do its job.
The exception: if you need a new paragraph, say "new line" or "new paragraph." That creates the visual break.
Add uncommon words to the custom dictionary
If you work in tech, medicine, law, or any field with specialized vocabulary, you'll say the same jargon repeatedly. And the transcription engine will get it wrong repeatedly.
Don't fix it manually every time. Add it to the custom dictionary once.
I added "PostgreSQL," "Kubernetes," "OAuth," and about 20 other technical terms. Now they transcribe perfectly every time.
See the custom dictionary guide for step-by-step instructions.
Speak at normal volume and pace
I thought louder and slower would be more accurate. I was wrong.
Speaking unnaturally slow confuses the engine because it expects normal speech rhythm. And yelling creates distortion that makes words harder to parse.
Speak like you're having a conversation with a friend sitting across from you. Natural volume, natural pace.
The only exception: if you're in a very noisy environment, you might need to speak slightly louder so your voice cuts through the background. But even then, don't shout.
Hold the phone correctly
I used to hold my phone right up to my mouth like I was whispering a secret. That created muffled, distorted audio.
The sweet spot: 15-25cm from your mouth (about half a forearm's length). Point the bottom microphone toward you.
Too close: you get popping sounds on hard consonants (P, B, T) Too far: background noise competes with your voice Just right: clear signal, natural tone
See mic positioning guide for detailed tips on dealing with wind and background noise.
Use wired EarPods when walking
If you're recording while walking outside, wired EarPods give you cleaner transcripts than AirPods.
I tested this for a week straight. Outside with AirPods: average of 6-7 errors per 100 words. Outside with wired EarPods: 1-2 errors per 100 words.
The difference is Bluetooth. AirPods constantly negotiate the signal in noisy environments, which can blur consonants. Wired EarPods don't negotiate. The signal stays consistent.
Full comparison: wired vs wireless guide.
Test and adjust
Everyone's voice is different. Your accent, your speaking style, your environment — these all affect accuracy.
Try this one-time test:
- Pick a paragraph you know well
- Record it in Brain Dump
- Compare the transcript to the original
- Note which words got wrong
Then adjust:
- If names are wrong → add to custom dictionary
- If sentences run together → pause more between sentences
- If technical terms are wrong → add to custom dictionary
- If lots of [inaudible] → check mic distance and background noise
Do this test once a month. Your accuracy will keep improving.
The 80/20 rule for voice notes
You don't need perfect transcripts. You need transcripts good enough to use.
If you apply just three of these techniques:
- Start with a title
- Use short sentences with pauses
- Add common jargon to the custom dictionary
You'll get 80% of the accuracy improvement. The rest is fine-tuning for specific situations.
When to give up and type
Some things are genuinely better typed:
- Long URLs
- Complex technical code
- Heavily formatted tables
- Numbered lists with specific formatting
For these, capture the core idea by voice, then open the note in a text editor and add the technical details by keyboard.
Voice is for thinking and capturing. Typing is for precision formatting.
Related guides:
- Custom dictionary setup for adding jargon and names
- Mic positioning for outdoor and noisy environments
- Markdown from voice for creating structure while speaking