Launched this week
Speech To Markdown
Harness local AI for notes
61 followers
Harness local AI for notes
61 followers
A free macOS and iOS app that turns your voice into structured markdown using local LLM. Global dictation anywhere with ⌘⌥], live markdown editor with on-the-fly fix with LLM. No cloud, no API keys — nothing leaves your Mac.
Unotone
Out of all the prototypes I've built to leverage local LLMs, the one I actually ended up using daily is my Speech-to-Markdown (STMD). Your speech goes through Local Whisper (STT) ➡️ gets processed by a local LLM ➡️ and is finally converted into a structured spec ✅
Using Fable, I finally managed to turn it from a web app into a native Mac app 💻, and I've been using it even more since. Whenever I have a new idea, I just open the app and start talking. My product ideas backlog has never been longer; now I just need more time (and more tokens 🤑).
🔗 https://voice-to-md.xajik0.workers.dev/ | https://github.com/xajik/voice-to-md
🎙️ The very first idea I "talked" into it was a Dictation mode for STMD — finally giving a purpose to the custom buttons on my keyboard ⌨️. Just focus on any text field, press a hotkey, and start talking. You can use it with or without an LLM to automatically fix grammar and typos.
📱 I also published an iOS version leveraging Apple Intelligence, but somehow it didn't click (https://apps.apple.com/us/app/speechtomarkdown/id6789775244 🍎). I guess most of my ideas come while I'm already sitting at my laptop.
Fully local is the right tradeoff for notes-to-self — those are exactly the thoughts you don't want leaving the machine.
The part I'm curious about is the markdown side rather than the speech side. Spoken thinking is rambling and non-linear; markdown is structured. So something has to decide that this sentence is a heading and those three are a list, and I'd guess getting that wrong is more annoying than a plain transcription error. How much structure does it try to infer on its own versus waiting for me to say it out loud?
Local Whisper plus a local LLM into structured markdown is a strong privacy story for people who draft long notes or specs. The global hotkey for dictation in any text field is what I would actually use day to day. Curious which local model you recommend for heading and list structure without butchering technical terms, and whether users can pin a preferred outline style.
fully local is the right call for something that's basically your unfiltered notes-to-self. curious what the hardware floor looks like though - does it bundle a specific small model sized for even an older/base-tier Mac, or does quality noticeably scale up on the higher-end Apple Silicon chips? asking because "local LLM" quality can vary a lot more by device than a cloud API would.
on-device is a smart default for something this personal. one thing I'd want to know before switching my note workflow over - does it handle mixed languages mid-recording ok, or does accuracy drop hard if you switch languages mid sentence like a lot of bilingual people actually talk?
looks useful. i'd also spend more time on the design. right now it feels a bit too ai-generated. i'd make it feel more human and give it a stronger visual identity.
also, why would i switch from wispr flow, superwhisper, or macwhisper? i'd make that value prop much clearer on the landing page.