Clips timed on https://talkhawk.app/dictation-on-a-cpu
Synthesized with Windows System.Speech, default voice, 16 kHz, 16-bit, mono.

== short, 5.2 s ==
Hey, can you send me the deck before Thursday? Thanks a lot.

== medium, 21.0 s ==
Quick update on the launch. The build is ready for testing, and I fixed the crash when you paste into Outlook. Please try it on your machine today and tell me what breaks. If everything looks fine, we ship on Thursday morning. I will write the release notes tonight and send them to the whole team before lunch tomorrow.

== minute, 51.7 s ==
I spent most of the morning on the onboarding flow. The first screen now explains the shortcut in one sentence, and the second screen asks for the microphone only when you are ready to try it. We removed the account step completely, because nobody needs an account to dictate. The trial starts the first time you open the app, and it runs for two weeks. During that time every feature is on. When the trial ends, dictation keeps working, and only the rewriting features need a plan. I tested the whole flow on three machines this afternoon. On the oldest laptop the model took about four seconds to load, which feels slow, so I want to load it in the background while the window opens. That should hide most of the wait. Tomorrow I will look at the settings page and the way we show errors.

== fivemin, 4 min 20 s ==
I spent most of the morning on the onboarding flow. The first screen now explains the shortcut in one sentence, and the second screen asks for the microphone only when you are ready to try it. We removed the account step completely, because nobody needs an account to dictate. The trial starts the first time you open the app, and it runs for two weeks. During that time every feature is on. When the trial ends, dictation keeps working, and only the rewriting features need a plan. I tested the whole flow on three machines this afternoon. On the oldest laptop the model took about four seconds to load, which feels slow, so I want to load it in the background while the window opens. That should hide most of the wait. Tomorrow I will look at the settings page and the way we show errors. The support inbox had a few interesting messages this week. One person writes long emails in German and switches to English in the middle of a sentence. The app handled it better than I expected, but the punctuation around the switch was wrong twice. Another person uses a foot pedal instead of the keyboard shortcut, which I had never considered. The pedal sends a key combination, so it works if you record that combination as your shortcut. A third message asked whether the audio is stored anywhere. It is not. The recording stays in memory until the text is ready, and then it is gone. I replied to all three and added the pedal setup to the help page. Here is the plan for the next release. First, the history page needs search, because people keep more than fifty dictations and scrolling is painful. Second, the indicator should show a clear message when the microphone is muted, since several people thought the app was broken when it was only muted. Third, I want to reduce the memory use while the app sits in the tray. An option to unload the model after an hour of silence would help people on smaller machines. None of these are large, so I expect the release to be ready in about two weeks if nothing surprising comes up. I tried dictating this whole update instead of typing it. The experience is strange at first, because you have to think in full sentences before you speak. After ten minutes it becomes natural, and I noticed that I write shorter, clearer sentences when I say them out loud. The tricky parts are names and numbers. Numbers come out well when you say them plainly, like twenty four or three point five. The biggest benefit is speed. I can say in one minute what would take me three or four minutes to type, and my wrists are happier at the end of the day. Before I forget, a few notes about the server. The rewriting service had two short slow periods on Tuesday, both under ten minutes. In both cases the provider was slow, not our code, and the app fell back to the plain text as designed. Nobody lost a dictation. I added a counter so we can see how often the fallback happens, without storing any text. The monthly bill for the model is still tiny, so cost is not a concern yet. That means one message when the error rate stays high for fifteen minutes, and nothing for single failures. I will set that up on Friday. I spent most of the morning on the onboarding flow. The first screen now explains the shortcut in one sentence, and the second screen asks for the microphone only when you are ready to try it. We removed the account step completely, because nobody needs an account to dictate. The trial starts the first time you open the app, and it runs for two weeks. During that time every feature is on. When the trial ends, dictation keeps working, and only the rewriting features need a plan. I tested the whole flow on three machines this afternoon. On the oldest laptop the model took about four seconds to load, which feels slow, so I want to load it in the background while the window opens. That should hide most of the wait. Tomorrow I will look at the settings page and the way we show errors.