Ben has made some recordings in the past weeks on the busiest places in the Bernouilliborg on the University campus. In stead of looking for speech or other foreground signals, the goal was to find loud background signals.
Imagine a robot trying to understand you inside a cafeteria or during the break at a lecture. Our human microphones – our ears – are perfectly capable of removing the background ‘mumbles’ of other conversations in order to listen to the sound source they want: their conversation partner.
Preliminary results show that – compared with a directional microphone – the word recognition rate increases five fold when recognizing spoken commands. Where the robot used to understand only 12% correct, now it’s about 72%. There’s still room for some optimization, so that even our robot can hear the forest through the trees.