Seven days. Three releases. Six blog posts. One Google Ads campaign that secretly added 19 countries without telling us.

Here is what actually happened this week building Superscribe.

Smart Transcription: The Big One

The headline feature was Smart Transcription, and it touched every layer of the stack.

The idea is simple: when you stop talking, Superscribe cleans up what you said. Removes filler words. Formats the output into whatever structure you need. Email, meeting notes, code comments, invoice line items.

The implementation was not simple.

We built a two-tier filler word system. The first tier is a fast regex pass that strips common fillers across five languages: English, Estonian, German, Spanish, and French. No API calls, no latency. Just pattern matching.

The second tier sends the transcript through GPT-4o-mini for intelligent filtering with a 4-second timeout. We tried Gemini 2.5 Flash first but reverted after it kept timing out in production.

On top of filtering, there is now a template system. Pick a template before you start talking. When you stop, the raw transcript gets transformed into structured output. Meeting notes come out as bullet points with action items. Emails come out ready to send.

One subtle detail that matters more than it sounds: the 750-millisecond stop buffer. When you release the hotkey, the mic stays open for another 750ms. This captures the trailing words people naturally speak as they reach for the keyboard. Without it, the last word or two would get cut off every time.

This shipped as v0.2.29 on February 24, both macOS and Windows.

Windows Parity Got Real

Making Superscribe work identically on macOS and Windows is one of those problems that sounds easy and absolutely is not.

The v0.2.32 release fixed a chain of Windows-specific bugs:

The filler filtering was silently broken because the GraphQL schema was missing an ID field on UserSettings. Apollo cache updates need a stable ID to merge data correctly. Without it, the mutation would succeed on the server but the local cache would never update. The toggle looked like it worked, but the setting never stuck.

Templates had a similar issue. The app was sending the string “none” instead of null when no template was selected. The backend treated “none” as a valid template ID, tried to look it up, failed, and silently returned unformatted text.

The most interesting fix was the non-activating window. On macOS, there is a clean AppKit API for windows that accept clicks without stealing focus from other apps. On Windows, you need Win32 interop. We used P/Invoke to call into user32.dll and set WS_EX_NOACTIVATE and WS_EX_TOOLWINDOW on the extended window styles. This means Superscribe can float over your editor, show streaming text, and never steal your cursor position.

Then v0.2.33 fixed a subtle bug where ambient mic noise could trigger the template processor. If you started a recording, selected a template, then stopped without actually speaking, the formatter would wait forever for structured output that would never arrive. The fix was a single line: trim whitespace from the final text before checking if it is empty.

Google Ads: The 145 Surprise Recommendations

We sat down to review the Google Ads campaign and discovered something alarming. Between February 16 and 22, Google had auto-applied 145 recommendations to our campaign. This included expanding our targeting from 4 countries to 23 countries without any human approval.

The good news: some of those auto-expansions actually worked. The campaign now reaches 27 countries and the overall numbers are solid. 12,214 impressions, 1,476 clicks, 12% click-through rate, 75 conversions at an average CPA of 8.10 euros.

The interesting discovery was geographic performance. Baltic and Southern European countries massively outperform English-speaking markets. Latvia converts at 2.91 euros per acquisition. The United States costs 16.03 euros. We added UAE, New Zealand, Hong Kong, and Norway based on the data.

We also disabled auto-apply for future recommendations. Trust but verify does not work when Google is spending your money.

Demo Mode: Try Before You Sign In

We designed and partially built a demo mode that lets people try Superscribe without creating an account. The backend is complete: 5 minutes of free transcription per device, identified by a UUID install ID, with IP rate limiting to prevent abuse.

The macOS client work is designed but not yet shipped. When it goes live, new users will be able to hit the keyboard shortcut and start dictating within 30 seconds of installing the app. No email, no password, no friction.

Build Infrastructure

Small improvements that compound over time. The Windows build script now creates missing directories before syncing to the VM. The installer reads version numbers dynamically instead of hardcoded strings. A new –build-only flag skips the S3 upload during development. A standalone signing script lives on the Windows VM for when SSH output gets truncated.

Every release goes through the same pipeline: build, sign (Apple notarization on Mac, Azure Trusted Signing on Windows), upload to S3 with versioned and latest paths, deploy the API, push to git.

Six Blog Posts

We published six blog posts on a one-per-day drip schedule. An SEO comparison piece against SuperWhisper. Three articles on dictation workflows and time tracking. A “voice memo graveyard” article about why voice notes die in your phone. And a meta piece about the prompting bottleneck.

Each post got a unique generated image and we fixed a date handling bug that was causing future-dated posts to appear before their scheduled time.

The Numbers

By the end of the week:

• 3 releases shipped (v0.2.29, v0.2.32, v0.2.33)
• 6 blog posts published
• 39 directory submissions total
• 27 countries targeted in Google Ads
• 75 conversions at 8.10 euros CPA
• Smart Transcription live on both platforms
• Demo mode backend complete

Next week: shipping the macOS demo mode client, more blog content, and continuing the directory submission campaign.

Try Superscribe free for 5 minutes, no signup required: superscribe.io

Related reading

• Fixing Estonian Dictation and Windows Streaming
• Smart Transcription: What Happens After You Stop Talking
• Apple Dictation Fails Multilingual Users