wtf is steno?

Steno is a Windows app for live transcription of calls that keeps track of who said what without guessing. It runs entirely on your own machine using whisper.cpp for speech recognition, so no audio is sent anywhere else. While a call is happening it listens to two separate audio streams at once, your microphone and whatever sound the system is playing, labeling one as you and the other as the remote party. Because the two streams are captured separately rather than mixed together, the app never has to guess who is speaking, unlike tools that rely on a voice matching model to sort out different speakers. It works with any app that produces sound, including Teams, Zoom, Discord, Google Meet, or a browser, since it captures system audio directly rather than plugging into each app individually. A related smaller app called Steno Dictate uses the same underlying engine but for a different purpose: you speak, and the words get typed directly into whatever program you have selected, correcting itself in place as you talk rather than waiting until you finish a sentence. Installing either app does not require administrator rights or a separate .NET installation, since everything needed is bundled together, and the two apps share the same core libraries so installing both barely costs any extra space. On first launch, the app downloads its speech recognition model, with a medium sized option around 1.5 gigabytes, and offers a smaller, faster model for weaker machines. The app also includes a built in speed test that measures how quickly your own hardware can transcribe a sample sentence and picks the best available option automatically, since integrated graphics chips can sometimes perform worse than the processor itself. Speed varies a lot depending on hardware. On a discrete graphics card the app can show a word appearing about seven tenths of a second after it is spoken, while the same task on a sixteen core processor with no graphics acceleration can take around eleven seconds per phrase, and the app tells you which mode is running at all times. During a call, two level meters and matching colors for each speaker make it easy to notice a muted microphone or misrouted audio, and there is a setting to reduce your microphone picking up the other person's voice through your speakers, though the README notes this is only a rough fix, not true cancellation.