Understanding And Visualizing Speech Quality
summary
A semi‑supervised encoder produces a dense embedding that predicts a 0–10 quality score, per‑segment scores and diarized per‑speaker quality, and classifies degradation types; training uses a few hours of human‑rated data plus tens to hundreds of hours of simulated degradations (experiments at ~40–400 hours) and the model cuts error by ~33% vs prior methods. The platform exposes diagnose/analyze/speech APIs as asynchronous file‑based endpoints (beta), returns JSON with loudness, noise, worst segment and speaker metrics, and shows degradation clusters in the latent space.
AI-written summary · about
chapters (6)
- 0:00Introduction to speech quality
- 8:11Algorithms behind quality scoring
- 24:04APIs for media diagnostics
- 49:40Q&A on quality metrics
- 1:06:30Visualizing audio analytics
- 1:13:50Closing and feedback
related talks
21mCan AI Disrupt Speech Compression?
17mContent Aware Encoding for low latency live streaming encoders using deep learning
18mBlind Quality Assessment for Real-Time Visual Communications
06mSoundLip: EnablingWord and Sentence-level Lip Interaction for Smart Devices
13mSENSEI: Aligning Video Streaming Quality with Dynamic User Sensitivity
09mCMSD and the Temple of Doom: Escaping VQA's Legacy Traps
20mLive Q&A 3 with Speakers
26mObjectionable Uses of Objective Quality Metrics