Live Q&A 3 with Speakers
summary
Live Q&A covers concrete systems and results: test infrastructure measures audio and video quality via a dedicated media‑quality team and plans to plug those tools into the new infra; Newton uses libwebrtc fake clock to run hundreds to thousands of parallel simulations, aggregate metrics and reproduce outliers; Opus work adds optional ML models with runtime/size tradeoffs and SDK releases; overlap speech separation, on‑device ML only during loss segments, and Kubernetes/AWS scaling limits with on‑prem as mitigation.
AI-written summary · about
chapters (16)
- 0:00Live Q&A introduction
- 0:25Audio/video quality metrics
- 1:30Open sourcing RTC work
- 2:34Adapting to Meta’s infra
- 3:16Upgrading Opus with ML
- 4:00Handling metric opinion variation
- 5:04VMAF vs no-reference models
- 6:51Artistic bias in scoring
- 8:16Newton and LibWebRTC clock
- 9:20Realistic test configurations
- 11:26Simulations for optimization
- 12:50Training datasets for bias
- 14:15Mobile complexity and performance
- 16:44Overlapping speech in transcription
- 18:07Scaling challenges and solutions
- 19:29Closing remarks
