Talat’s AI meeting notes stay on your machine, not in the cloud

The AI-driven notetaking application, Talat, is making waves as a private, local-only alternative to existing services like Granola, which has garnered a valuation of $250 million. Developed by Yorkshire-based programmer Nick Payne, Talat is designed for users who prefer to keep their data secure on their own devices rather than in the cloud. Payne, who describes himself as a computer enthusiast, was inspired to create Talat after exploring the capabilities of Granola. He found it remarkable that Granola could capture audio on his Mac without recording video—an impressive feat at the time. This curiosity led him to a new and lesser-known Apple API, prompting him to develop an open-source audio library named AudioTee to enhance his work with audio streams. Despite his admiration for the hosted transcription models used by competitors, Payne was troubled by the need to share not just his text data, but also his voice recordings. His breakthrough came when he discovered FluidAudio, a Swift framework that permits local, low-latency audio processing on Apple devices, enabling transcription models to run directly on the Mac’s Neural Engine. The culmination of Payne's exploration and development efforts resulted in Talat, a compact 20 MB application available for a one-time fee. Users can enjoy its features without needing to create an account or share any analytics. Talat captures audio during meetings on platforms like Zoom, Teams, and Meet, transcribing it in real time while also providing options to manage speaker assignments and take additional notes. At the conclusion of meetings, Talat generates a local summary that encapsulates key points, decisions, and action items, all of which are searchable within the app. Talat emphasizes user control over data handling, allowing users to choose their own language model (LLM), export options to applications like Obsidian, and utilize standardized connections for external data sources. The AI behind Talat is multifaceted, primarily leveraging FluidAudio. For summarization, it defaults to the Qwen3-4B-4bit model, which is efficient enough to run on modest hardware. Users also have the flexibility to utilize various cloud LLM providers or Nvidia’s speech recognition models, enhancing their personalization of the experience. At launch, the app is available for free with a trial of 10 hours of recordings for M-series Mac users. Talat is priced at $49 during its pre-release phase, with plans to increase to $99 upon full launch. Both Payne and his colleague Mike Franklin are committed to maintaining Talat as a one-time purchase product, ensuring users can take advantage of its features without ongoing costs.