FluidAudio

3 instead of the native Apple API and (b) do audio decoding and inference in a pipelined fashion. These two changes effectively eliminated the overhead. mpg123 is quite a bit faster than the native Apple API at mp3 decoding (has some very optimized arm64 assembly routines), and the pipelining ensures that the inference is never starved by the mp3 decoding. Contributing this pipelined setup to FluidAudio would be good.

apps are built around https://github.com/FluidInference/FluidAudio, if anyone is curious. Their readme has a big list of similar apps as well.

easier to merge segments or reassign a speaker than it is to untangle an incorrect merge. https://github.com/MimicScribe/benchmarks/blob/main/diarizat... The app provides in-meeting talking points using a prompt tuned for discovery type calls. It can suggest probing questions to help you extract more detail or helps you refocus on the big picture with “magic wand” type questions (e.g. “how would your ideal system work”). Getting low latency models to provide novel, relevant, and totally not hallucinated information is a bit of a reach and it tends to restate the transcript frequently but little gems do come from it sometimes so it’s best to think of it as a source of inspiration and be a vigilant gatekeeper. It’s set up so recording can be started and ended via holding a keyboard shortcut instead of connecting to your calendar service. I prefer this for privacy and to keep transcript history from getting cluttered. Tapping the shortcut shows and hides an always-on-top overlay on your active screen regardless of whether you have other apps full-screen or not. Beyond simple navigation, you can also use voice commands to make post-meeting corrections or additions, for instance, you can simply say "merge this speaker with that speaker" to clean up the transcript. It also has push-to-talk/dictate functionality with LLM cleanup - what the app started as but that tool was developer catnip, soo many of them. A developer friend who’s worked in finance reviewed the site and said he’d bounce because the privacy story wasn’t strong enough so I added a completely on-device mode and a bring-your-own-key option. Using cloud models does add a lot to the experience, including context aware speaker merging and fragment cleanup, summary items during meetings, action items attributed, etc. On-device mode is completely free and the speaker identification is still very useful. The privacy story is my biggest worry with the app, particularly since its target audience is more technical people. I’d love to get people's thoughts on it and any feedback would be super helpful.

frontier cloud LLM, and your summaries live on Granola's servers. There are lots of upsides to this: the transcription speed and quality are unmatched, the summarisation is awesome, and your meeting notes sync across all your devices, though all at the cost of privacy. I've been building up my toolkit in this space for a while: System Audio recording, custom notification windows, Acoustic Echo Cancellation, automatic meeting detection - different parts of the puzzle I've slowly pieced together over the past year or so. Recently discovering FluidAudio[1] finally let me build something I've wanted for a while: an app which privately transcribes your meetings in realtime, captures both sides of each call, and stores and summarises entirely on device. It's called talat[2], and I launched an early pre-release version last week. It's a Tauri app, which means the download size weighs in at 20MB. It's also a commercial project - I'm self-employed and need to make a living - but I hope even in the day and age of AI-commoditised software, there's enough in the app today and the direction of travel to justify the one-time purchase cost. While default-private is important, so is empowerment, so it lets you configure almost everything, including whether you want to use a different on-device or cloud LLM for summarisation. Lots of stuff needs improvement. Speaker diarisation is pretty rough. The app is macOS only and needs an M-series chip due to its use of the ANE for transcription. Local LLM summarisation (Qwen3-4B-4bit is the default) can be hit and miss, though in truth the more I use talat, the less I care about summarisation and the more I care about the transcript simply being there, ready to refer back to later as and when I want to. I'd really value your feedback, and I'd be delighted to answer any questions. The implementation was hugely accelerated through Claude Code, but the architecture, the design patterns, and all the little hard-won pieces of the puzzle mentioned earlier are entirely human. You don't need an account to use it, and there are no in-app analytics tracking what you're up to. [1] https://github.com/FluidInference/FluidAudio [2] https://talat.app

n the PyTorch and pyannote code, but the initial benchmarks look promising. We’d love feedback! We're currently working on adding VAD and integrating Parakeet for transcription, but still wrestling with CoreML model conversion.