Talat: the AI app that protects your data locally

Talat: A Local Alternative to Granola

In the realm of AI-powered note-taking applications, Granola has quickly established itself as a benchmark, reaching a valuation of $250 million. This app is particularly favored by tech startup founders and venture capital investors. However, a developer identified an opportunity for a more private and local solution, without a recurring subscription. Thus, Talat, a new application for Mac, was born.

Nick Payne's Initiative

Nick Payne, a passionate developer from Yorkshire, England, decided to take on this challenge. Describing himself as a "computer nerd," Payne explained that the idea of creating a local AI note-taker emerged from a series of fortunate coincidences. "Granola is a fantastic app, a brilliant example of what can be accomplished with Electron," he told TechCrunch. His curiosity was piqued when he discovered that Granola could record audio on a Mac without capturing video, a feat that led him to explore a little-documented Apple API.

Development of the AudioTee Library

To better work with this API, known as Core Audio Taps, Payne developed an open-source audio library called AudioTee. This initiative allowed him to gather various tools, although he did not immediately find a marketable product. The hosted transcription models used by applications like Granola are impressive, but Payne was concerned about having to share his personal audio data.

The FluidAudio Tool and the Neural Engine

The discovery of FluidAudio, a Swift framework enabling low-latency local audio AI on Apple devices, was pivotal. This framework allowed him to run transcription models directly on the Mac's Neural Engine, Apple's dedicated hardware for AI processing. This opened the door to creating a product where audio remains on the user's device, without being sent to external servers.

Designing Talat with Mike Franklin

Talat was developed in collaboration with Mike Franklin, a long-time friend and former colleague of Payne. The result is a lightweight 20 MB application, available for a one-time purchase, without requiring account creation or data sharing. Although Talat offers fewer features than some competitors, it focuses on the essentials: capturing and transcribing audio in real-time during meetings on platforms like Zoom, Teams, and Meet.

Features and Options of Talat

Talat allows users to reassign speakers, take notes, and edit transcriptions. The application attempts to assign speakers in real-time, but you can reassign them if necessary. You can also edit, delete, or split segments of transcription. At the end of each meeting, a local language model generates a summary of key points, decisions, and actions to be taken. For the summary part, the application defaults to an AI model called Qwen3-4B-4bit, which can operate even on relatively modest hardware. All data is stored locally and is easily searchable.

Customization and Future Integrations

Payne emphasizes the importance of configurability, offering users the choice of their language model, automatic export to applications like Obsidian, and the use of webhooks for data retrieval. The application primarily uses FluidAudio for its AI functions, but users can opt for other models like Qwen3-4B-4bit or Nvidia's Parakeet. Additionally, Talat allows users to replace the default AI model with any cloud LLM provider of their choice or use the Ollama tool to run models locally. An MCP server is mentioned as a standardized way for AI tools to connect to external data sources. Over time, Talat plans to add support for more built-in choices as well as integrations for other applications, such as Google Calendar and Notion.

Availability and Pricing

Currently, Talat is available to Mac users equipped with Apple M-series processors. They can test the application for free for 10 hours before deciding on a purchase. In its preliminary version, Talat is offered at $49, but the price will rise to $99 upon the release of version 1.0. Payne and Franklin are self-funding the project, planning to maintain the one-time purchase model in the future.