Dots Lab · experiment 003 · desktop tool
The dictation app worked.
Just not on our voice.
Glass Desk is a voice tool built for an accent the speech models were not trained on. Tap the key, talk, tap it again, and the words land in the window you were already typing in. It keeps a record of what it mishears, so it gets better at one specific voice instead of an average one.
It is free. One person dictating all day does not reach the free tier of the model host it runs on.
Hover, and it opens. The shortcut is the only label it needs.
01 The problem
It heard someone else.
Speech models are tuned on the accents they were trained on, and ours is not one of them. The fast variant returned Spice Nemo for voice memo and PyTLow for voice flow. Nothing downstream repairs a transcript that wrong, so the first decision was to give the speed back and run the full size model.
That fixed the words it had a chance at. It did not fix the names. Tool names, brand names and product names are exactly the words a general model has never heard us say, and they are most of what we dictate.
- we said
- voice memo
- it wrote
- Spice Nemo
- we said
- voice flow
- it wrote
- PyTLow
02 The run
One run, start to finish.
A full run at its real speed, nothing sped up and nothing cut. The pill in the film is the same component as the one above, rendered frame by frame from the app's own stylesheet, and the waveform is a real recording's levels. The sentence is written for the page: what people actually dictate into the tool stays on their own machine.
03 The pipeline
Two models, doing two different jobs.
One hears. One reads it back and repairs it. Splitting them is the whole trick, because the thing that is good at transcribing audio is not the thing that knows what Higgsfield is.
- 01
Hear whisper-large-v3
The full size model, never the turbo variant. Turbo trades accuracy for speed, and an accent is exactly where that shows: it returned Spice Nemo for voice memo. Roughly seven cents an hour more buys a transcript worth repairing.
- 02
Look up deterministic
Known mishearings are replaced in code, not in the prompt. Listing the pairs for the model leaks: told that price memo means voice memo, it also rewrote the price of Deepgram. Exact phrase replacement cannot generalise, so it is not asked to.
- 03
Repair llama-3.3-70b-versatile
A second model reads the transcript back with a glossary of names the first one mangles. Temperature zero, a non reasoning model on purpose: the reasoning ones spent the whole token budget thinking and returned nothing.
- 04
Insert no review step
The text goes to the clipboard and a paste is synthesized into whatever window had focus. No confirm keypress, no review pane. A dictation tool you have to approve is a text box with extra steps.
04 The interface
One object. Every state. No colour.
This is the pill at real size, running the app's own stylesheet and waveform code. Hover it, click the dot to record and click again to stop, or step through the states.
The pointer arrives and the sliver opens into a 186 by 40 control. Ctrl+Shift+Space is the only label.
State is weight, never colour
Recording is a filled dark square, the universal stop mark, not a red light. Nothing on the screen can clash with whatever is behind it.
At rest it is four pixels
Fifty two by four, parked on the bottom edge. Light rather than dark, because a dark sliver on a dark editor is invisible.
The window never resizes
Windows floors a frameless window at 64 pixels, so a four pixel sliver cannot be a window at all. The window stays a fixed, oversized, transparent 560 by 132, and the pill grows from 52px to 470px inside it, entirely in CSS.
No backdrop blur
Windows paints its blur to the window rectangle and ignores the rounded region, so real frosted glass would have forced square corners. These are near opaque surfaces instead, and the corners stay round.
05 What it learned
It keeps the mistakes.
Every dictation is filed with both versions, the raw transcript and the one that was inserted, plus the audio. That record is the product. A tool that hears one accent well cannot be shipped pre trained, it has to accumulate.
The same repair, five ways. What the first model returned, and what the second one made of it.
A repair is not the same as a lesson. Whether a pair is ever applied on its own is decided by the two gates below, and most are repaired from scratch every time.
Learning is gated twice
A new pair is only kept if the corrected side hits a real glossary term and the two sides actually sound alike. Anchoring on the glossary alone let Spice Nemo become ElevenLabs, because a guessing model reaches for the glossary exactly when the input is garbage.
And it still misses
Only multi word phrases apply themselves. Single words wait for a human, because the moment remote was learned as memo, remote team would have become memo team. The honest state of this tool is that it is better on names and still wrong on rare ones.
06 What it costs
It is free, and not the trial kind.
No seat, no plan, no word cap, because there is nothing to sell you. It runs on your own key at the model host, and one person dictating all day stays inside the free tier. Months of real use, nothing paid. No trial that expires and no word counter ticking down in the corner.
It still keeps a meter, because a tool that can spend money should say so. Every dictation's length and token use goes into the local record, so on the day you do push past the free tier you see the number as it happens instead of finding it on an invoice. For reference, past that line the full size model is about eleven cents an hour of audio.
| What is compared | Glass Desk | A dictation app you subscribe to | The one built into the OS | A meeting transcriber |
|---|---|---|---|---|
| What you pay | Nothing. One person's dictation stays inside the model host's free tier. Past that it is the model's own rate, cents an hour. | A monthly seat. Free tiers usually cap the words or the minutes. | Nothing. | A monthly seat, usually priced per person. |
| Where the audio goes | Your disk, and whichever model host you pointed it at. | Typically the vendor's cloud. | Depends on the setting. Some of it stays on the device. | The vendor's cloud, because the whole product is the archive. |
| Your accent | Every raw transcript is kept, and the names it mangles are repaired against a glossary of words you actually say. | Usually a custom words list you maintain by hand. | A general model. Yours is one of the accents it averages. | A general model. |
| Where the words end up | At the caret, in the window that already had focus. | At the caret. | At the caret. | In a document you then copy out of. |
| The recordings | Stay on your machine as labelled training data you own. | Held by the vendor under their retention policy. | Usually discarded. | Held by the vendor. That is the point of it. |
| The model | Open weights and swappable. Whisper large v3 and Llama 3.3 70B today, something better tomorrow. | Whatever the vendor ships this month. | Whatever the OS ships. | Whatever the vendor ships. |
This compares shapes of tool, not brands. Every product draws these lines differently and prices move, so check the one you use rather than trusting a table. The Glass Desk column is what it has actually cost to run every day since it was built, which is nothing.