Why online transcription services violate NDAs
Lawyers, journalists, HR managers, and podcasters routinely record meetings, interviews, and depositions that contain privileged information. When you use services like Otter.ai ($15/mo), Sonix ($22/mo), or even OpenAI Whisper API, your audio is uploaded to remote servers for processing. Even if these companies promise encryption, your recording physically exists on a server you don't control.
For any organization bound by NDAs, attorney-client privilege, HIPAA, or GDPR, this is a material breach. The solution is obvious: run the AI model on the client's own hardware so the audio never leaves the device.
How Whisper.js Works in Your Browser
- ONNX Runtime: The Whisper model is converted to ONNX format and executed via ONNX Runtime Web, which leverages your GPU (via WebGPU/WebGL) or CPU.
- Model Caching: The ~40MB model is downloaded once from Hugging Face and cached in your browser's IndexedDB. Subsequent uses load instantly.
- Audio Processing: Your audio is decoded via the Web Audio API, resampled to 16kHz mono, and processed in 30-second chunks — all in browser memory.
Frequently Asked Questions (FAQ)
Is my audio uploaded to any server?
No. The Whisper AI model runs entirely in your browser via Transformers.js and ONNX Runtime. Your audio is processed locally on your GPU/CPU.
Is this NDA-safe for legal recordings?
Yes. Since all processing happens locally on your device, no data is transmitted. Unlike Otter.ai or Sonix, there is zero risk of your confidential meeting recordings being stored on third-party servers.
How accurate is the transcription?
This tool uses OpenAI's Whisper model which achieves near-human accuracy for English and many other languages. The browser version uses the 'tiny' model variant for fast performance, which may be slightly less accurate than the full model.