Granola, an AI-driven note-taking application, with an estimated worth of $250 million, has become a widely used instrument among entrepreneurs and venture capitalists in the technology sphere. Yet, a particular developer contends there’s a market for a more confidential, exclusively local substitute, offered for a single payment and requiring no ongoing subscription. This conviction led to the creation of a new Mac application named Talat.
Nick Payne, a developer based in Yorkshire, England, who describes himself as a computer enthusiast, explains that the concept to construct a local AI note-taker emerged primarily from a sequence of fortunate coincidences.
“I consider Granola remarkable; it stands as a brilliant demonstration of what can be achieved with an Electron app [a framework for developing desktop applications] when given ample attention and refinement,” he informed TechCrunch. “Upon my initial trial, I was intrigued that it managed to record system audio on my Mac without capturing video, which was the customary workaround at that juncture. This sparked extensive investigation, leading to the discovery of a relatively new and sparsely documented Apple API.”
To simplify interaction with that API (Core Audio Taps, which allows developers to access a Mac’s audio streams), Payne decided to forge an open-source audio library, AudioTee.
“During that period, I was progressively assembling a suite of tools, but I never encountered anything that felt capable of existing independently as a product, rather than merely an impressive technical demonstration,” Payne commented. “The cutting-edge hosted transcription models — the very same providers utilized by entities like Granola — are extraordinary, and witnessing your speech unfold on-screen in near real-time is viscerally cool. Nonetheless, it consistently bothered me that the compromise necessitated supplying not just my data, but my audio data; my actual voice,” he elaborated.
He then chanced upon a software framework dubbed FluidAudio, a Swift framework that facilitates entirely local, low-latency audio AI on Apple devices. It allows the execution of compact, rapid transcription models directly on the Mac’s Neural Engine — Apple’s dedicated hardware for AI processing.
That was the crucial element that prompted Payne to realize he could transform his research into a tangible offering — one where your audio never departs your Mac, and your transcripts are not stored on the servers of another corporation.
Talat, developed in collaboration with Payne’s long-standing companion and former associate, Mike Franklin, is the culmination of Payne’s fascination with the audio domain. The outcome is a 20 MB application available for a single acquisition, which neither necessitates creating an account nor sharing analytical information with the creators. Furthermore, there are no continuous charges.
While some AI note-takers might possess a broader array of functionalities, Talat provides a refined set of capabilities. It captures sound from your computer’s microphone when you are engaged in conferencing applications such as Zoom, Teams, Meet, and others, and converts it into text instantaneously. The application attempts to assign speakers in real-time, but you retain the option to reassign them as required. You can also jot down observations, in addition to modifying, deleting, or dividing transcript segments. Upon the conclusion of the meeting, a localized LLM generates a synopsis detailing key points, decisions, and action items.
The notations, transcripts, and summaries are all searchable within Talat, too.
Beyond the aspect of privacy, Payne stated that the objective is to furnish users with more alternatives.
“We are emphasizing customizability and empowering users to dictate the destination of their information: select your preferred LLM, automatically export to [note-taking app] Obsidian, utilize webhooks that transmit data upon a meeting’s completion, an MCP server,” which is a standardized method for AI tools to link to external data sources, “to retrieve it on demand,” he clarified.
Beneath its surface, the AI comprises a blend — “primarily interconnected and abstracted behind FluidAudio,” Payne remarked, crediting it with undertaking a significant portion of the intricate work. For the summarization component, the application defaults to an AI model named Qwen3-4B-4bit, which is capable of operating even on relatively modest computing power.
Nevertheless, users have the flexibility to swap that out for any cloud LLM provider of their preference, or they can select between two Parakeet variations — speech recognition models devised by Nvidia — or direct it towards Ollama (a utility for running AI models locally), thereby granting them greater command over the experience. In due course, Talat will incorporate support for additional integrated choices, as well as integrations for other applications, such as Google Calendar and Notion.
At its debut, users with M-series Mac computers (those equipped with Apple’s proprietary processors, commencing with the M1) can download the software and test it for no charge with 10 hours of recordings before opting to acquire it.
Talat is procurable for $49 during this pre-release iteration, which remains under active enhancement.
When the application reaches its 1.0 version, the cost will escalate to $99.
Payne and Franklin are self-funding Talat and intend to maintain the core offering as a one-time acquisition going forward.
{content}
Source: {feed_title}