Microphone required for dictation.
Requires HTTPS or localhost
THE SERVICE, IN-BROWSER SPEECH RECOGNITION ENGINES, AUDIO WORKLETS, DICTATION WORKSPACES, AND ALL TRANSCRIBED TEXT OR GENERATED TRANSCRIPTS ARE PROVIDED “AS IS” AND “AS AVAILABLE”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE IMPLIED WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, TITLE, AND NONINFRINGEMENT.
WebScribe and its maintainers make no representation, warranty, or guarantee regarding the accuracy, completeness, precision, punctuality, acoustic fidelity, or suitability of automatic speech recognition (ASR) outputs and transcribed text for any particular purpose. Automatic transcription is probabilistic and subject to ambient acoustic noise, microphone quality, speaker accent, vocabulary limitations, and model inferencing constraints.
Dictated transcripts must NEVER be relied upon for medical diagnosis, clinical care, emergency dispatch, courtroom reporting, legally binding agreements, or safety-critical applications without independent human transcription and verification. Users assume sole responsibility for reviewing, proofreading, and verifying all transcribed content prior to commercial, legal, or production use.
100% Client-Side In-Browser Processing: All microphone capture, audio preprocessing, neural network inference, and transcript generation execute 100% locally within your device’s browser memory using client-side WebAssembly SIMD and ONNX Runtime Web. Your voice, raw audio recordings, audio waveforms, and dictated text are strictly confined to your local device and are NEVER uploaded, streamed, transmitted, or stored on any remote server or external network.
Acoustic Model Delivery & Local Storage: When a speech recognition model is loaded, the pre-trained open-source model weights (e.g., Moonshine, Nemotron, Parakeet) are downloaded once from the server or public catalog and cached directly in your browser’s private local storage (CacheStorage / IndexedDB). Model inference thereafter runs entirely offline on your local CPU / WebGPU hardware without transmitting any audio or speech data back to the server.
Local Browser Workspace & Clipboard Isolation: The dictation workspace operates purely in browser memory. Actions executed via standard browser interfaces (such as the “Copy & Clear” clipboard mechanism) interact directly with your device’s native operating system clipboard and are neither intercepted, monitored, nor transmitted to any third party.
Strict Privacy & Zero Telemetry: WebScribe does not track users, collect analytics, set advertising cookies, or record dictation sessions. No personal identifiers, IP logs, or search/speech records are gathered, profiled, or shared with third parties or external data brokers.
Zero Account & Zero Cloud Storage Dependency: No user credentials, accounts, logins, or server-side profiles exist. WebScribe maintains zero remote databases of user audio, transcripts, or personal data. Your speech and dictated text exist solely in your active browser window.
Instant Purge & Local Storage Control: Clicking “Copy & Clear” instantly wipes the transcript from local browser memory. Any local neural model weights cached in browser CacheStorage / IndexedDB remain strictly under your local device control and can be purged at any time via the “Clear Cache” action in Settings.
Agentic Development: This application was developed agentically with autonomous AI coding agents assisting in its architecture, system design, implementation, and UI engineering.
Zero Generative AI / Deterministic Runtime Guarantee: WebScribe does NOT connect to, invoke, query, or integrate any Large Language Models (LLMs), generative chatbots, cloud AI agents, or commercial generative APIs (such as OpenAI ChatGPT, Anthropic Claude, or Google Gemini) during runtime operation. The speech recognition pipeline consists entirely of open-source acoustic transducer and neural ASR models (e.g., Moonshine, Parakeet, Nemotron) and Silero VAD running strictly as local, deterministic ONNX inference pipelines on your device without generative hallucination, text completion, or autonomous agent execution. User audio and transcripts are NEVER forwarded to external AI model providers or used to train public generative models.