Developer runs Whisper speech-to-text locally to keep sensitive work call audio private

A developer building a work call transcription tool has chosen to run OpenAI's Whisper model locally rather than use cloud APIs, citing privacy concerns over sensitive client and pricing discussions. The setup runs faster-whisper in int8 quantized mode on an RTX 3060 GPU with 12 GB of VRAM, shared with two other AI models on the same server. To manage limited VRAM, the three models run sequentially, and Whisper falls back to CPU processing when GPU pressure is high, slowing transcription from roughly 8–12 seconds to 30–60 seconds per audio minute. Transcription accuracy exceeds 90% on clean audio but degrades noticeably with background noise, speakerphones, or domain-specific terminology. The developer acknowledges that cloud-hosted Whisper endpoints are faster and easier to manage, but argues the tradeoff favors local processing when the audio content itself is confidential.