The success of a language experiment heavily relies on selecting appropriate stimulus materials. This selection process entails a critical trade-off between similarity to ‘real’ language (i.e. external validity) and experimental and analytic control (i.e. internal validity). In order to bridge these conflicting demands, we developed the NyU-BU contextually controlled stories Corpus (NUBUC) of spoken language. The corpus is both naturalistic and experimentally controlled, comprising 16 high-quality recordings of 8 unique stories, spoken both by a female and a male actor. Each story consists of 128 sentences (~2000 words per story) organized around critical keywords, which have been matched along multiple linguistic dimensions. The context surrounding each keyword is also parametrically manipulated, varying prior context (weak/strong), local context (weak/strong) and sentence position (early/late). Here we describe the corpus in detail, including how it compares to and builds on existent corpora. These materials showcase the ability to overcome the apparent dichotomy between control and generalizability, by presenting subjects with carefully curated linguistic materials in a naturalistic listening scenario.

* These authors contributed equally This work was funded by a research grant from the USA Air Force Office of Scientific Research (AFOSR) awarded to OG [grant number FA9550-18-1-0055], NYU Abu Dhabi Institute Grant G1001 (LG) and the William Orr Dingwall Foundation (LG). Code which was used to process the corpus can be found here: https://github.com/polvanrijn/NUBUC