Your keyboard is the bottleneck, not the modelYour keyboard is the bottleneck, not the model

There's a version of this problem that everybody notices, and a version almost nobody does.

The one everybody notices: typing is slow. You know this. You've felt it in the gap between having a thought and finishing the sentence that contains it.

The one nobody notices: typing is slow enough that it changes what you say. And when the thing you're typing to is an AI, that changes what you get back.

That second problem is what this post is about.

The quiet cost of a keyboard

Think about the last time you asked an AI to build something for you. A component, a script, a feature, a landing page, whatever it was.

Now think about how much you actually told it.

If you're like most people, you typed three or four sentences. You described the goal, maybe a constraint or two, and hit enter. Then the AI gave you something in the general neighbourhood of what you wanted, and you spent the next twenty minutes correcting it. No, not like that. I meant this. Also it needs to handle this case. And the styling should match the rest of the app.

Every one of those corrections was context you had in your head the entire time. You just didn't type it.

Why not? Not because you forgot. Because typing costs effort, and your brain runs a silent cost-benefit calculation on every sentence before you commit to it. Is this worth typing out? Probably not, I'll just fix it after.

That calculation is the bottleneck. Not the model. Not the prompt engineering. The friction between your brain and the text field.

Speaking is three to four times faster than typing

Here's the part that makes this concrete.

The average person types somewhere around 40 words per minute. Trained typists push higher, 60 or 70 or 80, but most of us, typing normally with thinking mixed in, land in that 40 range.

The average person speaks at 120 to 140 words per minute in normal conversation. Not rushed. Not performing. Just talking.

That's three to four times the throughput, using an interface you've been training on since you were two years old.

But raw speed is only half the story. The more interesting number is the one you can't measure directly: how much more you're willing to say when saying it is cheap.

When a sentence costs you eight seconds of typing, you skip it. When it costs you two seconds of talking, you say it. Multiply that across every sentence you didn't bother with, and you get the difference between a prompt that describes what you want and a prompt that actually specifies it.

Context is the whole game

Everyone working with AI tools eventually arrives at the same conclusion: the quality of what you get out is almost entirely determined by the quality of the context you put in.

Not the model. Not the wording. Not some magic phrase you found in a prompt library. Context.

The AI cannot read your codebase's conventions unless you describe them. It cannot know that you already tried the obvious approach and it broke. It cannot guess that this component needs to work inside a modal, or that your team hates that particular library, or that you're optimising for readability over cleverness because someone else has to maintain this.

All of that lives in your head. All of it is one sentence away from being useful. And all of it gets dropped, every single time, because typing it out felt like too much work in the moment.

This is why dictation matters more for AI work than it does for almost anything else. It's not a typing replacement. It's a context unlock.

What actually changes when you start dictating

I switched to dictating my prompts, and the shift wasn't subtle.

I stopped editing my thoughts before having them. When typing, I'd compress an idea down to its shortest form before writing it. Speaking, I just say the whole thing, including the messy parts, the caveats, the "oh and also" additions that turn out to be the most important detail in the prompt.

I one-shot far more often. The back and forth I used to accept as normal (three or four rounds of correction) mostly disappeared. Not because the AI got smarter, but because I finally told it everything up front.

I think out loud instead of thinking silently and then transcribing. This one surprised me. Speaking a problem out loud is a genuinely different cognitive mode. You catch gaps in your own reasoning as you hear yourself say them.

The friction is gone. Long prompts stopped feeling expensive. That sounds minor. It isn't. It changed how much I'm willing to explain, which changed what I get back.

"But I'd have to clean up everything I say"

This is the objection that kept me from trying dictation for a long time, and it turns out to be mostly obsolete.

Old dictation was raw transcription. Every "um", every false start, every pause, dumped straight into your text field. You'd spend more time cleaning it than you saved.

Modern dictation tools run an AI pass over what you said. Silences get dropped. Filler words disappear. If you list things out loud, they come out as a formatted list. You end up with clean, structured text without ever having said the words "bullet point".

That's the difference between dictation as a novelty and dictation as an actual workflow.

The tool I use: FluidVoice

I use FluidVoice, and the main reason I'm mentioning it by name is that it's completely free and open source, which in this category is genuinely unusual.

The good dictation tools in this space are paid. Most offer a free tier capped around 2,000 words per week, which sounds reasonable until you actually use dictation daily and burn through it in one afternoon. Then you're paying a monthly subscription on top of the AI subscription you already have, just for the privilege of talking instead of typing.

FluidVoice sidesteps that entirely, and it does it by running the AI model locally on your own machine.

That has two consequences worth knowing.

The good one: nothing you say ever leaves your computer. No audio uploaded, no transcript sent to a server. If you dictate client information, private code, or anything you'd rather not hand to a third party, it stays on your machine. Most paid tools can't offer that, because processing your voice in the cloud is how they work.

The cost: about 3GB of RAM, running in the background. I've run it on a 24GB machine with no issues, and 16GB should be fine. Below that, the local model probably isn't the right choice, but FluidVoice also supports cloud models via your own API key, so you're not locked out.

One more limitation worth being honest about: the local model is built for English. If you dictate in another language, you'll want to connect a cloud model instead. Something like OpenRouter gives you pay-as-you-go access to multiple providers with one key, and even then you'll usually spend less than a paid dictation subscription costs.

The takeaway isn't "download this app"

The tool is replaceable. The insight isn't.

If you're working with AI every day, building, coding, writing, shipping, the amount of context you hand the model directly determines the quality of what comes back. And right now, for most people, the thing standing between their brain and that context is a keyboard.

So stop typing your thoughts. Say them. All of them, including the parts you'd normally skip. Let the tool clean up the mess.

Then watch how much less back and forth it takes to get to the thing you actually wanted in the first place.