Developing Machine Learning Based Speech Enhancement Models for Teams and Skype
summary
Microsoft replaces the traditional DSP noise suppressor with a deep learning model that performs 7X better than the previous solution, trained on synthetic mixtures from LibriVox, AudioSet, Freesound and DEMAND and validated with crowdsourced p.808/p.835 tests and a million-call A/B experiment. They release open datasets and test framework and host DNS challenges, and ship an optimized model plus an objective metric DNS-MOS with PCC=0.97 that cuts the perceived background-noise token by 31%.
AI-written summary · about
chapters (12)
- 0:00Speech enhancement at Microsoft
- 1:51Deep noise suppression need
- 3:38Software 1.0 vs 2.0 for ML
- 7:11Training datasets for noise suppression
- 9:18Crowdsourced speech quality with P.808
- 11:29Hosting DNS challenges
- 13:13Objective speech quality metrics
- 14:59Model complexity vs performance
- 16:23Integrating models into Teams/Skype
- 17:27Echo cancellation and packet loss
- 19:53Future speech enhancement work
- 20:56Summary
related talks
18mUsing Machine Learning to Enrich and Analyze Real-time Communications
35mScaling Background Noise Filtration for AI Voice Agents
06mSensing to Hear: Speech Enhancement for Mobile Devices Using Acoustic Signals
22mBuilding next-gen speech synthesis for Bookmate audiobook service
06mCountering Acoustic Adversarial Attacks in Microphone-equipped Smart Home Devices
21mMultimodal Open Source at Kyutai, From Online Demos to On-Device
29mEngineering Machine Learning Technologies
42mdata.bythebay.io: Shubho Sengupta,