You have a 45-minute meeting recording. You don't need a 12-page transcript. You need the five decisions that were made, the three action items assigned, and the one thing everyone disagreed about.

Cloud services like Otter.ai and Fireflies do this — they transcribe, then summarize. They're also listening to every word of your meeting, storing it on their servers, and processing it under terms of service that give them broad rights to use the data for product improvement.

For internal strategy meetings, HR conversations, legal discussions, or any recording with confidential content, that's a problem. Not a theoretical one — a practical one that compliance teams and privacy-conscious individuals think about constantly.

The same workflow runs locally. Audio goes in, key points come out, and the recording never leaves your machine.

Two Models, One Pipeline

This workflow chains two AI models in sequence.

Step 1: Speech-to-Text. The audio recording is transcribed by Parakeet, running on the Neural Engine. This produces a full text transcript — every word spoken, in order.

Step 2: Summarization. The transcript is passed to a language model (Llama, Qwen, etc.) with a summarization prompt. The LLM extracts the key points, decisions, action items — whatever you ask for.

The elegance is in the chaining. You don't manually copy the transcript and paste it into a chat window. The pipeline feeds the output of Step 1 directly into Step 2.

The ModelPiper Workflow

Load the Transcribe & Summarize template. It's pre-wired: Audio Capture → STT → LLM → Response.

Record directly or drag in an audio file. The STT engine transcribes it, then the LLM processes the transcript according to its system prompt. The default prompt extracts a structured summary, but you can customize it — ask for action items only, a bullet-point recap, a formal meeting minutes format, or whatever you need.

Customizing the Summary

The LLM's system prompt controls what comes out. Some useful variations:

Meeting minutes format: "Extract attendees, decisions made, action items with owners, and unresolved questions. Format as formal meeting minutes."

Key decisions only: "List only the decisions that were made in this meeting. Ignore discussion, small talk, and tangents. Be concise."

Client-ready summary: "Summarize this conversation for a client update email. Professional tone, focus on outcomes and next steps."

Technical review: "Extract all technical decisions, architecture choices, and implementation commitments. Flag any disagreements or unresolved technical questions."

You edit the system prompt directly in the LLM block. No code, no configuration files — it's a text field in the visual pipeline editor.

The Privacy Case

Meeting recordings are some of the most sensitive content in any organization. They contain unguarded opinions, salary discussions, strategic plans, personnel decisions, competitive intelligence, and casual remarks that were never meant to be documented.

Uploading them to a cloud transcription service means those recordings exist on third-party infrastructure. Local transcription and summarization means they don't. The audio file stays on your disk. The transcript exists in your app. The summary is on your screen. That's it.

Try It

Download ModelPiper, install ToolPiper, and load the Transcribe & Summarize template. Record something or drop in a file. Edit the system prompt to get the summary format you want.

Audio, transcript, and summary — all on your Mac. Nothing else involved.

This is part of a series on local-first AI workflows on macOS. Next up: Live Translation — speak in one language, hear the translation spoken back.