Define the closed boundary
Document ingestion, processing, storage, access, retention and deletion. Check telemetry, model downloads and automatic backup paths.
Include upload workstations, temporary conversion files, result delivery and backup storage when describing the local processing boundary.
Fit the model to hardware
File size is not GPU memory usage. Model weights and computation dominate memory, while long recordings can be processed sequentially with global timestamps.
Benchmark the selected model and precision on the available accelerator; recording size alone does not predict inference memory requirements.
Control every copy
Use access control, encrypted transfer and storage, isolated temporary directories, content-free audit logs and explicit retention for both primary and backup copies.
Apply retention and access policy to derived text and indexes as well as the original media, because each remains sensitive content.
We show the practical differences between models in Our open-model transcription test on long Russian recordings.
Evaluate quality separately
Privacy does not solve quiet speech or rare names. Pin the model and configuration and test genuine difficult recordings before routine use.
Use representative quiet, technical and multi-speaker recordings to prove recognition quality separately from the privacy architecture.