Ok Google, What Am I Doing? Acoustic Activity Recognition Bounded by Conversational Assistant ...
summary
An add-on using a Raspberry Pi and ReSpeaker 4-mic array records only when Google Home lights indicate interaction and feeds 64-bin log-mel spectrograms into a VGGish-based CNN pretrained on AudioSet; trained on data from 14 participants doing 19 activities, it reaches about 84.85% precision and 85.67% recall (≈84% F1) on 30-second clips and drops to ≈57% F1 when limited to the 2.3-second mid-interaction segment, while automatic location inference from predicted activities hits ≈95% F1.
AI-written summary · about
related talks
22mUbicoustics: Plug-and-Play Acoustic Activity Recognition
20mExploring the hidden potential of sound data
47mThe Digital Hearth: An In-the-Wild Automated Group Speech to Text System
34mGCP for Data Scientists
01h 14mExploring the hidden potential of sound data
15mGCP for Data Scientists
44mGoogle Assistant powered by Containers, Machine Learning and NET on Google Cloud
42mWelcome to VOICE: Opportunities and Challenges for Brands