WORKFLOW PERFORMANCE · 19 September 2026
Measuring session bandwidth in AI medical scribes
The microphone stops. The network does not necessarily stop with it. Audio uploads, final processing, note generation and application assets all contribute to the experience.
Start at browser launch. End at browser close.
Our metric includes the entire isolated Chromium session. That captures products that upload during recording and products that move more data at finalisation without selecting a favourable slice of time for either.
For the six-English-consultation subset (the main table now covers all completed tests), mean measured traffic is 10.26 MB for Hanah, 14.80 MB for Freed, 17.66 MB for Lyrebird, 14.50 MB for Preve, 28.19 MB for PatientNotes and 37.83 MB for Heidi.
These are observations of configured sessions, not promises about all visits. Browser background traffic is included. Profile caching, application changes and visit length can affect the result.
WebSockets count too
We use Chromium NetLog socket and UDP byte events. Counting those transport bytes includes WebSocket transfers without adding the same decrypted payload again. A missing log does not become a suspiciously attractive zero.
This is still not a bill from an internet provider: network packet headers and TCP retransmissions are outside the measure.
Less data does not prove faster performance
Lower transfer volume can matter on constrained connections, but this experiment does not directly measure performance under controlled bandwidth limits. Application latency, processing time and connection quality also matter. We report bytes and time separately.