Security Design Principle
Honest Processing Window
The only named plaintext moment in transcription
Every AI transcription service has this moment: the audio must be "heard" once to become text. We choose to be honest about it β define its exact boundaries and immediately destroy all traces when it ends.
Precise Definition
When audio enters an AegisMeetings transcription host, it is decrypted in memory and processed by the speech recognition engine. This period is called the Honest Processing Window.
The window lasts only as long as transcription itself requires. Once complete, the audio key and plaintext are immediately wiped from memory β not written to disk, not written to the database, leaving no recoverable trace.
After the window closes, even AegisMeetings cannot decrypt your audio again.
Event Timeline
Your browser encrypts and uploads
The audio file is encrypted on your device and uploaded directly to cloud storage using a time-limited credential (SAS Token). Audio bytes never pass through our web server.
Transcription host memory (the only plaintext moment)
The audio key is unsealed here; the audio is decrypted in memory and fed to the speech recognition engine. The transcription host runs with swap disabled, reducing the risk of plaintext being written to disk. Custom vocabulary is also briefly decrypted in memory as ASR hotwords during the same window.
Key and plaintext immediately destroyed
As soon as the transcription task completes, the audio key and plaintext are wiped from memory. The database never stored any key capable of decrypting the audio; the task message is consumed and gone.
Encrypted output stored to cloud
The transcript is encrypted with your personal key (held only on your device) and stored back to cloud storage. Our servers cannot read it.
Why don't other services say this?
Every AI transcription service has an Honest Processing Window β without exception. The physics of speech-to-text require the AI to "hear" the audio. The difference: some pretend it doesn't exist, or hide behind vague phrases like "end-to-end encrypted" while staying silent about the transcription step. We choose to be honest about where it is, how long it lasts, and what happens when it ends.
Most services
Audio may be retained indefinitely or used to train models. Terms like "end-to-end encryption" sidestep the plaintext transcription step. Decryption keys may persist in the database permanently.
AegisMeetings
Honest Processing Window boundaries are explicitly defined. The audio key is never written to the database and never persists. Destroyed on completion. After the window closes, even we cannot decrypt your audio β this is an architectural guarantee, not a promise.
Ask your current service these four questions
-
When my audio is being transcribed, which host and which memory region holds the plaintext?
-
After transcription, where is the key that decrypted my audio stored β database? disk?
-
How long after transcription is the plaintext audio actually wiped?
-
Is my audio used to train or improve AI models?
Technical Verification
Read the Full Security Architecture Whitepaper
Ready to try a different kind of AI meeting recorder?
Free to start, no credit card required. Your audio is encrypted on your device. After the Honest Processing Window closes, we cannot decrypt it again.
Try FreeThis page is derived from AegisMeetings Security Architecture Whitepaper Β§1 and Β§8. For full cryptographic architecture, key hierarchy, and trust boundary details, see the whitepaper.