Local-First Whisper Transcription: Accuracy and At-Rest Encryption in Ibis v1.0

Cloud-based audio transcription services require uploading recordings to remote servers, a workflow that is incompatible with research, journalism, and legal practice involving confidential audio. We describe the design and substantiation of Ibis, a Windows desktop application that performs all transcription locally using Whisper `large-v3-turbo` with int8 quantization (via faster-whisper and CTranslate2) and stores processed audio encrypted at rest using the age v1 file format. We benchmark Ibis's transcription accuracy on the standard LibriSpeech test sets and report 1.94% word-error rate on test-clean and 3.87% on test-other, within 0.27 percentage points of OpenAI's published full-precision Whisper-large-v3 baselines. We document the encryption architecture in full, including key management via OS-native credential stores, three customer-control affordances (in-app replay, in-app bulk export, and a standalone command-line recovery tool), and the failure modes inherent to local-only key management. The benchmark harness, recovery tool, and encryption module are released under the MIT License with full source available for independent verification.