The Microphone Is Not the Product: Why Dictation Apps Live or Die on Polish

I tested a dictation app that transcribed with 99% accuracy. It was still unusable.

Here's why: it gave me exactly what I said. Every "um," every false start, every run-on sentence. The words were perfect. The writing was garbage.

That's when I learned: transcription accuracy is table stakes. Polish is the product.

What most reviews measure (and why it's wrong)

Every dictation review focuses on one thing: word error rate. Did the app get the words right?

I used to think this mattered most. Then I started timing myself. Transcription accuracy was 96% on my last app. But I still spent 8 minutes editing every 5-minute dictation. Because accuracy isn't the same as usability.

The words were right. The writing was wrong. And editing writing takes ten times longer than editing typos.

The two jobs of dictation software

Most people think dictation is one job: turn speech into text. It's actually two.

Job 1: Transcription. Turn audio into words. This is hard, but it's solved. Apple's speech framework, Whisper, modern cloud APIs , they all do this well. 95-98% accuracy is standard now.

Job 2: Polish. Turn raw words into finished text. Fix the "ums," the false starts, the run-on sentences. Adjust tone. Make it sound like something you'd actually send.

Job 2 is where dictation apps live or die. And almost nobody talks about it.

The polish gap in the category

Here's what I noticed reviewing dictation apps: everyone claims to do polish. Almost nobody does it well.

Some apps just fix grammar and punctuation. That's not polish , that's spellcheck. The "ums" stay in. The run-ons stay in. You still sound like you're talking, not writing.

Other apps rewrite everything into fluent prose. But they lose your voice. The output sounds like AI wrote it, not like you polished it.

The good ones , and there are maybe two or three , thread the needle. They clean up disfluencies while keeping your voice. They adjust tone without erasing personality. They make you sound like a better version of yourself, not like a different person.

Echo Flow is my attempt at that third category. I've dictated thousands of sentences through it. It still sounds like me , just the version of me that slept well and had coffee.

Why this matters for the category

The dictation category is about to split into two camps:

Camp 1: Apps that transcribe well and stop there. They'll compete on accuracy numbers that don't matter to actual users.

Camp 2: Apps that transcribe well and polish well. They'll compete on something real , how little editing you have to do after.

I'm betting everything on Camp 2. Because I've been the person staring at a perfect transcript that I still can't send. And I know I'm not alone.

The bottom line

If you're evaluating dictation apps, don't just check the word error rate. Dictate something messy , with "ums," false starts, run-on sentences , and see what comes out.

If it's accurate but unusable, you'll spend more time editing than you saved by dictating. If it's accurate and polished, you've found your tool.

Polish isn't a feature. It's the product.

Try Echo Flow free for 14 days · See the rewrite palette · Why the same sentence should not land everywhere

Sources

• Apple SFSpeechRecognizer documentation: transcription accuracy on Apple Silicon Macs.
• Echo Flow's polishing pipeline: transcription → mumble-cleanup → AI polish → rule-based polish, all local.
• Writer's own testing: transcription accuracy vs. editing time across multiple apps.