Capture meetings and system audio, get an accurate transcript back, and let it draft the summary, decisions, and action items — all on your Apple Silicon GPU through local Whisper and Gemma models. No upload, no account, no cloud. Built for rooms where the conversation can't leave.
Silo Systems builds infrastructure for people who would rather not negotiate with their own software. Silo Transcribe is the same discipline, pointed at one job — turning a conversation into a searchable, summarized record — done entirely on the machine you already own. Your meetings never become someone else's training data.
Capture, transcription, diarization, and summarization all run on your Mac through a bundled local runtime on your Apple Silicon GPU. The speech and language model weights download once from public open-weights mirrors, then Silo Transcribe works offline for the lifetime of the app — in a locked room exactly the way it works at your desk.
No analytics SDK, no telemetry, no account, no crash reporter. The Python inference layer binds strictly to loopback (127.0.0.1) as on-device IPC — it is not a network service. The audio you record and the transcripts it produces stay in your Application Support container, never transmitted anywhere.
Structured summaries, decision registers, action items, and participant lists are generated locally by a Gemma instruction model through mlx-lm. The meeting that would normally be fed to a cloud notetaker never touches a server — the intelligence is computed on hardware you already own.
Silo Transcribe records the system audio of a call through ScreenCaptureKit and your microphone through AVAudioEngine at the same time, then merges them into one stream. Whoever is on the line and whoever is in the room end up in the same transcript — no virtual cable, no second device, no bot joining your meeting.
A local Whisper model (large-v3-turbo, MLX) transcribes the audio on your Apple Silicon GPU, with in-memory 16kHz mono resampling for speed. Energy-based diarization splits the result into per-speaker utterances, so the transcript reads like the conversation actually happened — not one undifferentiated block.
A local Gemma instruction model, running through mlx-lm, reads the transcript and produces a structured summary, a decision register, a list of action items, and the participants — all written to a local SQLite store you can search with full-text FTS5. The summary a cloud notetaker would charge you for, computed on your own machine and kept there.
Record system audio via ScreenCaptureKit and your microphone via AVAudioEngine at the same time — so a call, a recording, and the person in the room all land in one transcript.
Speech-to-text runs through a Metal-accelerated Whisper model (large-v3-turbo, MLX) with in-memory 16kHz mono resampling. Fast on Apple Silicon, and entirely on-device.
Energy-based partitioning splits the audio into clean per-speaker utterances, so the transcript reads like a conversation instead of a wall of text.
A local Gemma model via mlx-lm turns the transcript into a structured summary, a decision register, action items, and a participant list — no cloud notetaker required.
Every session is indexed with SQLite FTS5, so you can query across months of meetings instantly and jump straight to the line that matters.
Correct the transcript inline, retitle sessions, and export transcripts and summaries to a folder you choose. Your words, your files, no re-upload.
Whisper and Gemma download once from public open-weights mirrors with live byte-progress. After that, every transcription and summary is offline.
Start a capture from the menu bar, drop in an existing audio file to transcribe, and let smart titles name the session for you.
Cloud notetakers rent you transcription by the seat and run your meetings through their servers. Silo Transcribe is a single purchase, on your Mac, with unlimited local transcription and summarization — because it all runs on hardware you already own. Buy once, keep the record.
Coming to the Mac App Store ⌘Silo Transcribe is in pre-release for the Mac App Store. Join the waitlist and we'll send a single email when it goes live — no marketing list, no newsletter, no follow-up campaigns.
