Noted: Private, On-Device Meeting Transcription with AI Summaries and Speaker Identification

The Problem

Organizations routinely generate highly sensitive intellectual property, strategic planning, legal discourse, and patient data during internal meetings. Yet, the standard industry practice for recording and summarizing these sessions relies on cloud-based transcription services. These platforms require uploading raw, unencrypted audio to third-party servers, creating significant data privacy, security, and compliance risks.

For industries bound by HIPAA, GDPR, or strict corporate confidentiality agreements, sending audio to a remote server is often explicitly prohibited. Furthermore, cloud transcription services typically charge per minute or per seat, resulting in unpredictable and scaling operational costs.

Beyond security and cost, existing solutions frequently fail in challenging acoustic environments. In lecture halls, auditoriums, or large conference rooms, distant speakers and heavy background noise (such as HVAC systems or room reverberation) result in unusable transcripts. Existing tools rely on generic audio processing tuned for close-range headset microphones, failing to adapt to the physical realities of enterprise environments. Organizations require a solution that guarantees data sovereignty, predictable pricing, and high-fidelity transcription regardless of the acoustic environment.

Solution Overview

Noted is a native macOS application that brings the entire AI transcription and summarization pipeline directly to the user's device. By executing speech-to-text, speaker diarization, and Large Language Model (LLM) inference locally on Apple Silicon, Noted guarantees that zero audio or text data ever leaves the machine.

Designed to operate fully offline, Noted eliminates the attack vectors and compliance violations associated with cloud APIs. The architecture leverages Mistral's Voxtral and CoreML-optimized Whisper models for transcription, coupled with local LLM execution via the Apple MLX framework. This allows Noted to deliver real-time, highly accurate transcription, speaker identification, and structured meeting summaries without sacrificing privacy.

By processing everything on-device, Noted also introduces zero-latency dictation and advanced digital signal processing (DSP) capabilities that dynamically adapt to difficult acoustic environments, ensuring enterprise-grade reliability in any setting.

Key Capabilities

On-Device Transcription and Speaker Diarization

Noted captures system audio, microphone input, or a mixed composite of both, transcribing it in real-time. It supports automatic language detection and manual selection across 99+ languages. The system employs advanced speaker diarization to identify and label individual speakers within a conversation. Users can create named voice profiles by storing 256-dimensional speaker embeddings, allowing the system to automatically tag transcripts with specific names (e.g., "who spoke when") with precise attribution.

Local AI Summarization and Tagging

Once a recording concludes, Noted automatically generates concise session titles, 2-3 sentence structured summaries, and categorization tags. This is achieved via an embedded LLM server running directly on localhost. Users can choose from multiple optimized models, including Qwen 2.5 (1.5B/7B), Llama 3.2 (1B/3B), or DeepSeek R1, depending on their hardware capabilities. Because inference is local, users can summarize highly confidential board meetings or financial reviews without exposing the data to external AI providers.

Custom Fine-Tuned Model Integration

For specialized industries, generic transcription models often struggle with domain-specific jargon. Noted includes built-in support for custom fine-tuned CoreML models. Users can download specialized models directly from HuggingFace repositories or import them from local directories. This capability allows medical professionals, legal teams, and engineering departments to utilize models trained specifically on their industry's vocabulary, drastically reducing transcription errors on technical terms.

Large Room Acoustic Processing

To address the challenge of distant speakers in large rooms, Noted features a dedicated DSP audio preprocessing pipeline. When activated, this pipeline applies a high-pass filter to eliminate low-frequency rumble, FFT-based spectral noise reduction to strip out HVAC and background hum, and automatic gain control (AGC) to boost quiet, distant speech. This preprocessed audio is fed to the transcription engine while the raw audio is preserved for archival, ensuring high accuracy even in lecture halls or large boardrooms.

Screen and Multimedia Capture

Beyond audio, Noted integrates with macOS ScreenCaptureKit to record the user's primary display during sessions. It captures high-quality H.264 video with AAC audio at configurable frame rates and quality presets. It also supports capturing periodic screenshots at set intervals. All media is saved locally alongside the transcript, providing a complete, synchronized record of presentations, code reviews, or design workshops.

System-Wide Ephemeral Dictation

Noted includes a global dictation mode accessible via keyboard shortcut. Users can dictate text into any application on their Mac; the audio is transcribed in real-time and pasted instantly at the cursor position. This function operates ephemerally—no session is saved, ensuring a clean, frictionless workflow for drafting emails, documents, or code comments.

Architecture & Technical Design

Noted is engineered specifically for macOS, leveraging the platform's native frameworks to maximize performance and minimize battery consumption. The application is built using Swift and SwiftUI, ensuring a lightweight, responsive interface that resides unobtrusively in the menu bar.

The core AI workloads are accelerated using the Apple MLX framework and Metal GPU APIs. By compiling Metal shaders with explicit deployment targets and performing pre-flight hardware validation, Noted ensures stable GPU acceleration on Apple Silicon hardware, maximizing inference speed while managing memory pressure. The local LLM engine dynamically allocates resources, utilizing quantized models (ranging from 876 MB to 4.5 GB) that fit within the unified memory constraints of modern Mac hardware.

Audio capture relies on a robust implementation of AVAudioSession and ScreenCaptureKit, operating completely within the macOS application sandbox. The Large Room Mode preprocessing pipeline uses Apple's Accelerate framework (vDSP) for high-performance, zero-latency scalar and vector math, applying noise reduction and gain control without taxing the CPU.

Because the software requires no internet connection for core operations, its security posture is inherently absolute. Network access is only utilized for optional model downloads or software updates, meaning organizations can deploy Noted in air-gapped environments.

Use Cases & Scenarios

Corporate Legal Counsel

Challenge: A legal team frequently conducts internal strategy meetings and confidential client consultations. They require transcripts for record-keeping but are barred by strict non-disclosure agreements from uploading audio to third-party transcription services. Solution: The team deploys Noted on their enterprise Macs. They record the meetings using the application's local WhisperKit engine. The LLM generates structured summaries of the legal arguments and action items, while all data remains entirely on the physical device. Outcome: The legal team gains the productivity benefits of AI summarization without violating confidentiality obligations or risking IP leakage.

University Research Seminars

Challenge: A research department records guest lectures in large, reverberant auditoriums. Distant speakers and aggressive HVAC noise cause standard transcription tools to fail, producing garbled text. Solution: The researchers enable Noted's Large Room Mode. The audio preprocessing pipeline automatically strips out the low-frequency room rumble and applies automatic gain control to boost the speaker's voice before it hits the transcription engine. Outcome: The department achieves highly accurate transcripts of academic lectures, complete with AI-generated tags for easy searching and archiving in the session browser.

Medical Diagnostics Review

Challenge: Medical practitioners reviewing patient case files need to transcribe their notes quickly, but general-purpose transcription models fail to recognize complex pharmacological and anatomical terms. Solution: The clinic utilizes Noted's custom model support to load a fine-tuned medical Whisper model downloaded from HuggingFace. They use the system-wide dictation feature to speak directly into their Electronic Health Record (EHR) software. Outcome: Medical terms are transcribed accurately in real-time. The ephemeral dictation ensures that no patient data is permanently stored in the transcription application, maintaining HIPAA compliance.

Software Engineering Syncs

Challenge: A distributed engineering team conducts daily sprint planning and architecture reviews via video conferencing. They need a way to document technical decisions and code shown on screen without a dedicated note-taker. Solution: A team lead uses Noted to capture both system audio (the remote team members) and microphone audio, alongside a screen recording of the shared architecture diagrams. Speaker diarization labels who suggested which technical changes. Outcome: The team receives a synchronized video, labeled transcript, and a bulleted summary of action items immediately after the call, streamlining the development lifecycle.

Pricing & Plans

The direct-download edition of Noted is a desktop application for macOS. Users can choose between flexible subscription plans or a lifetime license for the current major version. Enterprise deployments with custom requirements are also supported.

Plan	Price	Features Included
Monthly	€5 / month	On-Device Transcription, Speaker Diarization, AI Meeting Summaries, Dictation Mode, Screen Recording, 99+ Languages, All Future Updates
Yearly	€48 / year	All Monthly features (Save 20% annually)
Buy Once (Lifetime v1)	€65 one-time	All features, valid for the entire v1.x lifecycle
Enterprise	Custom	Unlimited usage, SSO/SAML support, volume licensing, MDM deployment support, dedicated account manager

Pricing Philosophy: We believe in flexible ownership. Users who prefer a predictable operational expense can subscribe monthly or yearly, while organizations looking to capitalize software assets can purchase a lifetime license for the current major version.

Frequently Asked Questions

Does Noted require an internet connection? No. Noted is designed to be privacy-first and operates 100% on-device. An internet connection is only required initially to download your preferred ML models or to check for software updates. Once models are downloaded, you can disable your network connection and the app will function fully.

Are Intel Macs supported? Intel Macs support basic speech-to-text transcription. However, advanced AI features, including LLM-generated summaries and titles, rely on the Apple MLX framework and require an Apple Silicon (M1 or newer) processor to function.

Can I use domain-specific transcription models? Yes. Noted supports loading custom fine-tuned CoreML Whisper models. You can add these by providing a HuggingFace repository ID or by pointing the application to a local folder containing the necessary model configuration files.

How does Large Room Mode affect transcription? Large Room Mode is an opt-in audio preprocessing pipeline that applies noise reduction, high-pass filtering, and automatic gain control to improve transcription in noisy or highly reverberant environments. The original raw audio is still saved alongside your session for archival purposes.

Why Amsterdam Technologies

At Amsterdam Technologies, we build software grounded in an engineering-first philosophy. We prioritize native performance, architectural elegance, and uncompromising data sovereignty. By harnessing the full power of modern silicon, we deliver AI capabilities directly to the edge, freeing organizations from the latency, recurring costs, and security vulnerabilities of the cloud. Noted represents our commitment to building professional-grade tools that respect user privacy by design.

• Website: https://noted.amsterdam-technologies.com
• Contact: support@amsterdam-technologies.com
• Company: amsterdam-technologies.com