A little over 30 years ago (in late 1993), I recorded a video of myself demonstrating “The Berkeley Restaurant Project (BeRP)“system. BeRP was an example of a very early conversational AI system. It was a mixed-initiative spoken dialog system built at the International Computer Science Institute (ICSI) in Berkeley, California, USA in the early 1990’s.

BeRP was capable of having spoken conversations to help users find restaurants in the city of Berkeley. The automatic speech recognition (ASR) system used a “hybrid” decoder in which a neural network generated posterior probabilities for the sounds of English. These probabilities were used as the “emission” probabilities in a hidden Markov model (HMM) based speech decoder.

You can read more about this hybrid neural network HMM approach in our NeurIPS (NIPS) 1990 paper: Connectionist Approaches to the Use of Markov Models for Speech Recognition.

More information about this system: https://web.stanford.edu/~jurafsky/icslp-red.pdf

You can also read more about this project and download the transcripts and audio that we collected.