The story behind talat.
Like a lot of people, I love Granola. The onboarding, the delighters, the way it just works. It was the first app I'd ever used that recorded system audio on Mac without recording video too, and that blew my mind. I wanted to understand how it did it, which sent me down a rabbit hole into Apple's Core Audio Taps API that I've yet to emerge from.
That rabbit hole led to me building a couple of open source libraries for system audio capture (Swift and Node.js), some fascinating freelance work in the audio and transcription space, and a slowly-assembled toolkit: system audio recording, mic capture, acoustic echo cancellation, automatic meeting detection, custom window notifications, and a whole lot more. Different parts of the puzzle, pieced together over a year or so, but none of them a product.
One thing that always nagged me about cloud-based transcription tools is the tradeoff. The speed and quality are unmatched, the summarisation is awesome, and your notes sync across all your devices. But to get all of that, everything passes through someone else's servers: every frame of audio from your mic and other speakers on your call, your transcript, your notes, your summaries. Not just your data, but your audio data; your actual voice. It's a lot of links in the chain, and unavoidably it's just not private.
Recently discovering FluidAudio, which runs small, fast transcription models directly on the Mac's Neural Engine, was the piece that brought everything together. Seeing your speech unfurled onscreen in near real-time, entirely on your own hardware, is viscerally cool. I suddenly had everything I needed to build what I'd wanted for a while: an app that privately transcribes both sides of your meetings in real time, entirely on device. Think Granola, but local. Your audio never leaves your computer and your transcripts live on your machine.
That app is talat. I'm building it with Mike, who I've known and worked with for the best part of 17 years. He's the most brilliant engineer (and friend) I've ever met, so I want to work with him whenever I can. We're both based in Yorkshire in England; we don't have VC backing or funding, we don't have servers to run or investors to satisfy. We're bootstrapping talat between the two of us.
All VC-backed SaaS businesses need to start making their money back eventually, and fair enough. But it means your data tends to be the product, or at best the collateral. Metrics are everything, so you can be fairly sure that every single action you take, even in a desktop application, is being tracked meticulously. And what happens when the terms change, the price goes up, or you just want to leave? Your data is on their servers; you're locked in. talat is the opposite: the app has zero analytics, no event or activity tracking, nothing. And because we don't have servers to run, if you like what you see after 10 hours of free recordings, you can buy a lifetime licence for a one-off price. We're self-employed and need to make a living, but when anyone can build an increasingly functional version of almost anything with a single prompt, we believe the way to succeed is to offer something so good that the economics of building your own just don't stack up. The way we do that is to build an awesome product, listen to our users, and keep improving; same as it ever was.
When we started building talat we thought we'd need all the bells and whistles: in-app notes, polished summaries, the works. But we're discovering very quickly that most people (including us) don't really add notes, and they rarely read back summaries. What they care about is knowing the transcript is there if they need it, which is reshaping how we think about the whole product.
We're positioning talat as a conduit rather than a destination; we want to get out of your way as quickly as possible. Everything works out of the box, but everything is configurable: your LLM provider, your summarisation prompts, where your data goes. Auto-export to Obsidian, webhooks that push data out when a meeting finishes, an MCP server so your AI tools can query your meeting history. In the future we want to open up the input side too; plug in your calendar, your contacts, context from other tools, so talat can do a better job of things like speaker identification. The more you give it, the smarter it gets, and none of it leaves your device. Privacy is where we started, but empowerment is where we're going.
Lots of stuff still needs work. Speaker diarisation is pretty rough. There are sharp edges everywhere, and we know it. But we're iterating daily and the direction feels right. If you want to see what we've been up to, the changelog has the details, and if something isn't working or you want to tell us what to build next, we genuinely want to hear it.
AI is constantly reshaping so much of what we all do every day, but I can't see us ever losing that most human thing: talking to each other. I hope that never changes.
Nick & Mike
nick@talat.app · mike@talat.app