This repository contains the most recent transcripts of the audio files contained in the Berkeley Restaurant Project (BeRP) corpus. When corrections come in, I will update this repository.

The Berkeley Restaurant Project (BeRP) was a testbed for a speech recognition system developed by the International Computer Science Institute (ICSI) in Berkeley, CA, USA, in the 1990's.

The BeRP system was designed to be an automated consultant whose domain of knowledge was restaurants in the city of Berkeley. The system served as a testbed for several research projects, including robust feature extraction, neural-net based phonetic likelihood estimation, automatic induction of multiple pronunciation lexicons, foreign accent detection and modeling, advanced language models, and lip-reading.

The BeRP corpus, collected at ICSI, currently contains approximately 8566 utterances, with about 1900 unique words, comprising about 7 hours of speech. The audio data was recorded with a Sennheiser close-talking microphone and was sampled at 16 kHz.

The entire ~500MB BeRP corpus (audio and transcripts) can be downloaded from:

http://www.icsi.berkeley.edu/ftp/pub/speech/wooters/berp.tgz

Each line in the transcript.txt file contains the transcript for a single audio file. The first token on each line is the base name of the corresponding audio file. So, for example, the line containing the transcript for the audio file 4F_1_0010.wav.gz looks like this:

4F_1_0010 i wanna spend about five dollars

See the file transcription_guidelines.txt for the guidelines that the human transcribers were supposed to follow when transcribing the audio data. If you find any issues, please let us know so that we can update the transcripts!

In our BeRP paper from 1994, and the follow-on paper from 1995, we mention "1389 hand-written context-free rules". While I could no longer locate that specific set of 1389 rules, I was able to locate a set of 1412 rules dated Aug 1995, which are included in this repo as:

amer_trn2456_1412rules_aug95.cfg

(Caveat: I have not verified that this cfg matches the transcripts in this repo.)

D. Jurafsky, C. Wooters, G. Tajchman, J. Segal, A. Stolcke, E. Fosler, and N. Morgan, "The Berkeley Restaurant Project", Proceedings ICSLP, 1994. PDF

C. Wooters, "Lexical Modeling in a Speaker Independent Speech Understanding System", PhD thesis, University of California, Berkeley, CA, 1993. Available as ICSI TR-93-068. PDF

A YouTube video shot in Nov. 1993 of Chuck Wooters using the BeRP system.

Jurafsky, Daniel, Chuck Wooters, Jonathan Segal, Andreas Stolcke, Eric Fosler, Gary Tajchaman, and Nelson Morgan. "Using a stochastic context-free grammar as a language model for speech recognition." In 1995 International Conference on Acoustics, Speech, and Signal Processing, vol. 1, pp. 189-192. IEEE, 1995. PDF

The file wordhist.txt contains a frequency count of all of the words in the BeRP transcripts. It is generated with this command:

cut -d" " -f2- transcript.txt | tr " " '\012' | sort | \
  uniq -c | sort -rn > wordhist.txt