← The work
Voice & robotics Experimental
RockBot
With RockBot, I’m exploring when a robot should listen, reply or stop talking, using a local speech pipeline for Danish and English.
Conversation in a room.
Microphone audio passes through speech recognition, a decision about whether to respond, local language models and speech synthesis. The pipeline runs on macOS.
The interesting question is whether a reply belongs in the conversation. RockBot tries to distinguish speech addressed to it from nearby conversation, follow an exchange, suppress its own speaker echo and stop when someone interrupts.
Danish speech is transcribed and translated for the conversation pipeline; English input is supported too. I’m interested in whether it responds at the right moment as well as what it says.