While cloud-dependent note-takers like Granola dominate the startup scene, a new contender is betting on edge inference. Talat, a Mac-native application developed by Nick Payne, processes meeting audio entirely on-device, leveraging Apple's Neural Engine rather than external APIs. This approach addresses growing privacy concerns among engineering teams handling sensitive architectural discussions.

Payne built Talat after exploring Apple's Core Audio Taps API, eventually releasing an open-source library, AudioTee, to handle system audio streams without video workarounds. The real breakthrough came with FluidAudio, a Swift framework enabling low-latency local transcription. This architecture eliminates the need to upload sensitive voice data to third-party servers, a significant compliance advantage for enterprise environments.

For ML engineers, the flexibility is the standout feature. The default summarization model is Qwen3-4B-4bit, quantized for modest hardware. However, users can swap in Nvidia Parakeet variants, connect to local Ollama instances, or route tasks to specific cloud providers. Integrations extend beyond simple exports; Talat supports webhooks and an MCP server, allowing AI tools to pull structured data on demand into systems like Obsidian or Notion.

The app targets M-series Macs, offering real-time speaker diarization and searchable transcripts without requiring user accounts or analytics tracking. Currently in pre-release, Talat costs $49 with a 10-hour trial, scheduled to rise to $99 upon version 1.0 launch. Payne and co-founder Mike Franklin are bootstrapping the project, prioritizing a one-time purchase model over recurring subscriptions.

This shift suggests a growing appetite for private, configurable AI tooling among technical users who prefer controlling their inference pipeline over convenience. As data governance tightens in 2026, local-first solutions may become the standard for engineering teams.

Source: TechCrunch