Closing Address
summary
An experiment runs cloud transcription services on podcast audio and finds cloud transcription defaults are useless: services need tuned modes, fail on technical conversations, accents and multi‑party audio, so published episodes still use human transcribers. Unconference threads report no consensus on data versioning — teams commonly use GitHub for metadata while leaving data files on file systems. Edge inference discussions focus on privacy by keeping models on device with differential privacy and federated approaches. Full‑stack ML engineers remain rare.
AI-written summary · about
chapters (11)
- 0:00Introduction and welcome
- 1:16About TWiML and podcast success
- 2:21Cloud transcription experiments
- 4:29Unconference wrap-up session
- 5:11Data management and versioning discussion
- 6:17ML infrastructure and open source
- 7:20Conference app and resources
- 8:00Inference at the edge potential
- 8:45Making ML teams more efficient
- 10:11Future of the conference
- 10:53Podcast opportunities and thanks
