Heard
I built Heard to keep speech around a Mac available as timestamped text for an agent. It can capture microphone and system audio, transcribe locally, and write a plain record that another tool can read. Summaries and follow-up decisions belong to the agent using that record.
A small interface to an ongoing stream
Start, pause, stop and follow are the core interactions. Pausing keeps the models loaded; stopping flushes pending speech. The output stays in an ordinary local file instead of disappearing into a separate meeting-notes application.
I also added application-level audio exclusions. Excluded applications are filtered before their audio reaches speech detection or transcription. If the capture filter cannot be updated safely, system-audio capture pauses rather than continuing with an uncertain filter.
That separation makes the tool useful as a building block: capture and transcription have a clear job, while the user and their agent decide when to listen, what to retain and what to do with it.