RTC @Scale, Audio ML Session
summary
Live Q&A describes deployed deep noise suppression used in Skype/Teams, running on client endpoints with measured CPU cost of 1–2% CPU on Intel and a DSP fallback so >99% of calls use the ML path. Evaluation uses ACR for overall quality and CCR to catch dropped syllables. Codec work discusses Soundstream low-rate ~4–5 kbps trade-offs, FEC/redundancy for robustness, and a move from first‑gen speech‑only models to autoencoder-based codecs that are more fail‑safe.
AI-written summary · about
related talks
20mLive Q&A 3 with Speakers
18mUsing Machine Learning to Enrich and Analyze Real-time Communications
35mScaling Background Noise Filtration for AI Voice Agents
13mHow We Reduced Latency by 70% While Maintaining 99% Accuracy
09mUsing AI to get relevant KPIs for Audio Quality
23mContact Center Voice AI: Low-Latency Intelligence Extraction from Messy Audio Streams
15mOptimizing inference for voice models in production
19mAudio Streaming - Using WebRTC for building your own Voice AIs