Hi PH 👋 I built this because every meeting tool I tried had the same shape: to get a transcript, upload your meeting to someone's server. For a lot of conversations that is simply not an option, and "we take privacy seriously" is not an answer. EchoAI runs the whole pipeline on your machine. Speech recognition, speaker labelling, segmentation — all local. The only time anything leaves is if you explicitly ask a language model to tidy up the finished text, and you can skip that. What changed since the last version: - **macOS support.** It was Windows-only. Apple Silicon runs on Metal — real-time factor went from 2.04 on CPU (falling behind twice over) to 0.35 on GPU. Audio routing sets itself up, including installing the virtual device: one password prompt, no Homebrew, no Audio MIDI Setup, no manual aggregate devices. - **Five languages, detected per sentence.** Mandarin, Cantonese, English, Japanese, Korean — including switching mid-sentence, which is how people in Hong Kong and Singapore actually talk. This is the part I am most pleased with, and the part other tools handle worst. - **No key, no account, nothing to sign up for.** It transcribes out of the box. An API key is optional and only buys cleanup and reply suggestions. - **Speaker labelling**, with the voice prints kept, so if it over-splits one person into three you can tell it the real headcount at export and it re-groups after the fact. - **Pick the turns you want answered.** Cmd-click any set of turns in the transcript and ask for a reply to just those — which matters in a multi-party call, where answering everything is noise. - **Crash-safe.** Every finished sentence hits disk as it is spoken. An 84-minute meeting produced 1311 lines and used to ride entirely on the process staying alive. Now a crash costs the last sentence. - **A dead microphone recovers itself.** Bluetooth headsets stop delivering audio without saying so; it now detects that the callback has stopped — which is what separates a dead device from a muted one — and rebuilds the stream. Honest about the state of it, all of this is in the README: - Speaker labels are a guide, not a verdict. The voice print depends on *what* is said, not only who says it: the same person's speech scored against their own read-out digits comes to 0.473, below the threshold that decides "same speaker". No threshold fixes it, so the export lets you set the real headcount and re-cluster instead of pretending. - Wear headphones. On speakers your microphone hears the far end and files it under your own name — meeting apps cancel echo on the stream *they* send, which never touches a microphone opened separately. - The Windows capture path installs but I have not been able to test it on real hardware since reworking segmentation. MIT, no account, no telemetry, no paid tier. github.com/colakang/echoai_helper

GitHub - Live meeting notes that never leave your laptop

Replies