Learn AI with Zain Qurashi  ·  Day one lab, free  ·  Version 1 tonight
DRIP.

Day one lab  ·  Level 0  ·  free

Build your own dictation tool. Then make it worth $15 a month.

Hold a key, talk, get text. The first version takes about an hour, and honestly, anyone can ask an AI for one. What people pay for is everything stacked on top: the right models, running on your own computer, a style for every app, your phone, and a record of everything you've said. That's what you learn here, one layer at a time.

version 1 runs in Chrome or Edge. no python.

What the shops charge

Wispr Flow$15 / mo
Typeless$30 / mo
Version 1, tonight$0
Version 3, Scribe, in the Lab$0, forever
List prices on Oct 2, 2026; both drop to $12 on annual plans. They're good tools. The point is that you can build one, and learn how it works on the way.

Where it ends up  ·  an illustration

What it looks like a month in.

Not the bare version: the one you end up with. It hears you in any app, writes the way that app needs, works from your phone, and keeps a record of everything you've said, so you can see how you actually talk.

team chatday 31
#team chatstyle: casual

hold right ctrl
cleaning up...

Your notes, saved on this computer

  • Outline for the voice video: one model hears you, another one understands you.notes
  • Follow up with the client on the onboarding checklist.email
  • Split the sync job into a nightly batch and a webhook for new rows.code
●your phonestyle: notes

Talk into your phone. Your own computer at home does the work and sends the text back. No subscription, no cloud.

An illustration, not a live demo. Words dictated, pace, hours saved and sessions are my real numbers from Scribe; the words, phrases and topics are examples of what the insights look like. Yes, "go ahead and" is mine.

How it gets better  ·  layer by layer

Anyone can make version 1. Here's what you stack on top.

Ask any AI for a voice-to-text tool and you'll have one in minutes. It works, and it's bare. Each layer below is a lesson, and each one makes the same tool better. That's the difference between a toy and something you'd pay for.

1

Talk, get text

Version 1 · tonight

A button, your microphone, a speech service. Works in an hour. Your voice goes to someone else's server, and it's copy and paste from there.

2

Choose the model that hears you

Version 2 · free

Speech models come in sizes. Small ones are instant and miss names; big ones catch accents and jargon but want a graphics card. You'll test three on the same paragraph and pick yours.

3

Run it on your own computer

Version 2 · free

Download the model once and your voice never leaves the machine. Free to run, works offline, and a hotkey types straight into any app.

4

Add a model that understands you

Version 2 · free

Hearing words isn't understanding them. A language model fixes punctuation, drops the ums, turns "ten thirty" into 10:30 and keeps your meaning.

5

A style for every app

Version 3 · the Lab

Casual in chat, professional in email, word for word in your code editor. Plus shorthand: say a short phrase, get the full thing typed out.

6

Never lose the first word

Version 3 · the Lab

The mic stays open into a short memory buffer, so it rewinds a third of a second when you press the key. A small detail, and the one people notice most.

7

Your notes, your insights

Version 3 · the Lab

Every dictation saved on your computer and searchable. Then the fun part: your pace, the words you lean on, the phrases you repeat, what you talked about this week.

8

Your phone, too

Version 3 · the Lab

Put your phone and your computer on one private network with Tailscale. Talk into your phone; your computer at home does the work and sends the text back.

Models that hear

Speech to text. Bigger usually means more accurate and slower. Some are English only and very fast.

ExampleSpeedAccuracy
Whisper tinyruns on almost anything
Whisper large-v3-turbothe usual sweet spot
Whisper large-v3wants a graphics card
ParakeetEnglish only, very fast

Models that understand

The clean-up step. Same text in, very different results, costs and privacy.

  • A small model on your computerFree, private and quick. Great at punctuation and filler words, weaker at tricky rewrites.
  • A fast hosted open modelVery fast and cheap per word. Your text leaves the machine.
  • A big cloud modelBest at tone and hard edits. Costs tokens on every dictation, and your text leaves the machine.

The bars are rough, to show the trade-off. In the lab you measure your own, on your own computer, with the same 100-word paragraph.

Version 1  ·  tonight  ·  about an hour

Build version 1. The way everyone builds their first one.

You don't write code tonight. You describe the tool to an AI, save what it gives you as a file, open it, and paste any error back. That loop is the lesson. Use Claude, ChatGPT or Gemini; it doesn't matter which.

1

Describe it. Save it. Open it.

Copy this into your AI. Save the reply as dictate.html in a new folder. Open that folder in VS Code, install the free Live Server extension, and click Go Live at the bottom right. Your page opens at 127.0.0.1 in your browser; use Chrome or Edge.

Why not just double-click the file? Browsers only let speech recognition use your microphone on a real web address, and a file you double-clicked isn't one. 127.0.0.1 is a real address that lives on your own computer, so it counts.

Make me a single HTML file called dictate.html with no external libraries. A big button I can hold down (mouse, touch, or the space bar) to dictate using the browser's built-in speech recognition (webkitSpeechRecognition). While I hold it, show the words as they arrive. When I let go, put the final text in a text box with a Copy button and a Clear button. Keep a list of my last 20 dictations with timestamps in localStorage, with a button to clear it. Show a counter: words this session, and how many seconds the last dictation took from release to final text. If the browser doesn't support speech recognition, say so and suggest Chrome or Edge. Plain, readable code with comments, dark theme.

If it doesn't work, paste the error message (or "the button does nothing") back to the AI and try again. Two or three rounds is normal.

Want to see the finished one first? Try mine: it runs right here in your browser. Stuck after three rounds? Download it and compare it with yours. The comments explain every part.

2

Add a free key: a better model, and a clean-up step.

Make a free account at console.groq.com (no card) and create an API key. Then paste this to your AI. This is layers 2 and 4 in miniature: a better model to hear you, and a second model to understand you.

Add an optional key mode to dictate.html. A settings area where I paste a Groq API key, stored in localStorage, with a visible warning that this is only okay for a personal tool on my own computer. When a key is present: record audio with MediaRecorder while I hold the button, and on release send it as multipart form data (fields: file, model=whisper-large-v3-turbo) to https://api.groq.com/openai/v1/audio/transcriptions with the key as a Bearer token, then show the text. Add a "Clean up" button that sends the text to https://api.groq.com/openai/v1/chat/completions with a model chosen from a dropdown filled from GET https://api.groq.com/openai/v1/models (skip models whose id contains whisper or tts), asking it to fix punctuation and capitalization, remove filler words, keep my meaning, and return only the text. Show the tokens used from the response's usage field and keep a running total.

Now it works in any browser, the accuracy goes up, and you can see what clean-up costs in tokens.

3

Measure it. Keep the receipts.

Read the test paragraph below into your tool, count the wrong words, note the latency, and fill in the scorecard. You'll fill in the same scorecard for version 2 and version 3, and the difference will be yours.

  • Send me your numbers: leave your email at the bottom and reply to the welcome email. I read every one.
  • Keep the file. Version 2 starts from it.

What version 1 will be bad at

  • Browser mode needs Chrome or Edge, and sends your voice to Google's servers.
  • It clips the first word if you talk the instant you press.
  • No clean-up without a key: every "um" stays.
  • It's a web page, so the text doesn't land in other apps. You copy and paste.
  • Key mode is better, and still cloud: your voice goes to Groq.

What that's teaching you

  • What an API key is, and why it must never ship inside a product.
  • Why a memory buffer exists (so the first word survives).
  • Where your data goes, and what "runs on your computer" is worth.
  • What a clean-up model costs per thousand words.
  • That the first version always works and is always bare. The layers are the rest.

The scorecard  ·  fill it in before you leave

Your numbers, next to where it ends up.

The test paragraph · 100 words · read it in one go, no pause before the first word

Hi, this is a quick test of my dictation tool. My name is Maya Patel and I work with Jordan Reyes on the Tuesday report. We ship about 240 orders a week, and the API that tracks them goes quiet around noon. Yesterday I said the first word too quickly and the tool missed it. Numbers like 3.5 and names like Okonkwo are the hard part. If this comes back with every word, the punctuation in the right place, and nothing invented, the tool is better than most. Stop recording right after the last word, with no pause at all.
MeasureHowYour version 1Version 3: Scribe (mine, measured)
Wrong wordsout of the 100 in the paragraphmeasured in the Lab, on the same paragraph
Latencyseconds from release to text; your tool shows it0.92 s median
Cost to runper 1,000 words: the speech API's price plus clean-up tokens. Free tier = $0 today; write the list price$0, runs on your computer
Cost to buildtokens or dollars your AI used, from its usage screenabout 25,000 lines, with tests
Where the voice wentcloud, or your computeryour computer, nothing leaves it
First word survived?did "Hi" come throughyes: the mic rewinds a third of a second
Pacewords a minute while speaking160 wpm average, best session 245
Time saved so farwords spoken vs typing them at 40 wpm10.4 hours over 342 sessions
saved in this browser only. nobody sees it unless you post it.

Three versions of the same tool

Version 1 tonight. Version 3 is the one from the animation.

Each version fills in the same scorecard, so you watch your own numbers move. Version 2 is free. Version 3 is the app I use every day, and the Lab builds it with you.

Version 1 · tonight · free

The web page

  • One HTML file, built by describing it
  • Browser recognition, or Whisper on a free key
  • Your voice goes to the cloud
  • Clips the first word
  • Copy and paste into other apps
Version 2 · next · free

On your own computer

  • Layers 2 to 4: the model you picked, running locally, plus clean-up
  • About 300 lines of Python
  • A hotkey that types into any app
  • A lesson that shows your voice becoming numbers, then a spectrogram, then words
Send me version 2
Version 3 · the Lab

Scribe

  • Layers 5 to 8: styles per app, shorthand, the memory buffer, notes and insights, your phone
  • Speech to text on your GPU, a local model for clean-up: nothing leaves the computer
  • Members get the finished app the day they join, then build it
Unlock in the Lab

Why I built mine

Everything you say is data. It should be yours.

When you talk to a chatbot, your ideas live in its history, on its servers, in its format. Dictation is different: you're producing your own notes all day, in your own words. Kept on your computer, they're searchable, they show you what you keep coming back to, and you get better at saying it.

That's what my stats page is: 77,000 words in a month, which apps they went to, how fast I talk, how much time it saved. Mine, measured, on my machine. Version 1 starts that tonight: the prompt in step 1 keeps your last 20 dictations.

Scribe's stats screen: 77k words, 160 words a minute, 10.4 hours saved, a words-per-day chart.

Keep going  ·  free

Version 2 is free. I'll send it when it's ready.

Leave your email and you'll get version 2 (the one that runs on your own computer), the nine-level roadmap, and a note when the Lab opens. You don't have to buy anything to keep building this. Run wild with it.

free. version 2, the roadmap, then the odd next step. no spam, and you can leave from any email. privacy