Hacker News Reader

Parent Full thread

ben_jones

I'd imagine the priority might be:

1) Microphone all possible tables and have them record/signal when key words are overheard

2) Befriend/seduce/bribe wait staff and train them to roam around in optimal patterns to catch conversation

3) Some kind of subtle hearing amplification/focusing device

4) Record all conversations using a device on your person and process them later

View on HN

Record all conversations using a device on your person and process them later

Processing and extracting multiple conversations at different levels from a single audio source automatically would be a great project to attack with some quite simple sound engineering tools and speech recognition ML.

Cocktail party problem is actually hard in ML. And that's not even getting into speech recognition on imperfectly segmented speech streams.

Keywords: multiple talker speech segmentation, computational auditory scene analysis CASA

Reply on news.ycombinator.com