I Built a Meeting Notes App That Never Leaves Your Laptop
You know those meeting assistants that politely join your call as "Notetaker", record everyone, ship the audio off to some server and then send you a nice summary... and a monthly invoice? They work, sure. But every tim
You know those meeting assistants that politely join your call as "Notetaker", record everyone, ship the audio off to some server and then send you a nice summary... and a monthly invoice?
They work, sure. But every time one showed up in a call about a client contract or a project under NDA, I couldn't help wondering where that recording was going to end up... and I really needed something to be able to live translate any incoming voice, and fast (upcoming important work call).
So I built Parla ("speak" in Italian): a meeting notes app for Windows where everything, and I mean everything, happens on your own computer.
What it does
- Records any meeting, from any app (Zoom, Teams, Meet, a YouTube video, whatever). It captures your mic plus the system audio through WASAPI loopback, so no bot joins the call and you don't need virtual audio cables.
- Transcribes word for word, live, in 25 European languages.
- Tells speakers apart. You can rename them afterwards, or merge two labels when the model decides you're two different people (it happens).
- Translates live in a floating overlay, so you can follow the people speaking another language while they talk.
- Summarizes on demand: key points, decisions and action items, but only if you ask.
- Imports audio and video files, and exports everything to Markdown.
No accounts, no subscription, no cloud. The models download once during setup and after that it works offline.
The part I cared about most: modest hardware
A lot of local AI tools quietly assume you have a gaming GPU sitting under your desk. I wanted Parla to run on the kind of laptop most people actually have at work, so everything runs on the CPU and 8 GB of RAM is enough.
The first-run wizard checks how much RAM you have and only offers the summary model that fits: Qwen3.5 2B on 8 GB machines, with the option to switch to 4B if you have 16 GB or more. The translation model only gets loaded while translation is on, so it doesn't sit in memory for nothing.
Under the hood
The core is Rust, the UI is React inside Tauri, and the models are all open:
| Job | Model |
|---|---|
| Speech recognition | NVIDIA Parakeet TDT 0.6B v3 (int8) |
| Voice activity | Silero VAD |
| Speaker labels | WeSpeaker ResNet34 |
| Live translation | Tencent Hy-MT2 1.8B (Q4_K_M) |
| Summaries | Qwen3.5 2B / 4B (Q4_K_M) |
| LLM runtime | llama.cpp, CPU build |
Speech, VAD and speaker embeddings run through sherpa-onnx, while translation and summaries go through a local llama.cpp server. If you already have LM Studio or Ollama running, you can point the translate and summarize slots at those instead.
One fun problem: without headphones, your mic picks up the other people's voices from the speakers, and you end up with every sentence transcribed twice. There's an echo guard for that, which drops the mic's copy of what the system audio already caught.
Try it
Grab the installer from the releases page. It installs per user, so no admin rights needed, and everything it creates lives in one folder you can wipe from the settings.
It's free and open source (MIT), so feel free to poke around the code, open issues, or tell me what breaks on your machine.
Let me know if you like it, and happy (private) meetings!
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β full credit and traffic to the original publisher.