What is the switchboard corpus?

The Switchboard Corpus is a well-known dataset originally collected for government-funded research to advance technology in speech recognition. Established with support from DARPA, it consists of approximately 300 hours of recorded telephone conversations, involving speakers who discussed predefined topics of mutual interest. These recordings were initially captured using analog devices and later digitized. Each conversation lasted around ten minutes, and participants were paired based on shared topic preferences.

The transcription process involved creating a high-quality reference transcription by the Linguistic Data Consortium at the University of Pennsylvania. This reference transcription was used as a benchmark to assess the accuracy of speech recognition systems. Accuracy is measured through the word error rate, which calculates discrepancies like substitutions, deletions, and insertions between the machine-generated and the reference transcriptions. The Switchboard Corpus thus serves as a benchmark task for evaluating and improving speech recognition technologies 1.