Between Signal and Symbol: Sound, Speech, and the Data of Language

In 1969, J. R. Pierce, executive director at AT&T Bell Laboratories, called for a suspension of all speech recognition research, condemning the field as an “artful deceit” perpetrated by “untrustworthy engineers.” Automatic speech recognition, he insisted, could not be solved through engineering, and would be possible only once computers incorporated linguistic expertise comparable to a native speaker. Just two years later, IBM launched its Continuous Speech Recognition research group, which developed a data-centric approach that became standard not only in speech recognition and natural language processing, but across “big data” and machine learning applications for everything from financial modeling to bioinformatics. Frederick Jelinek, the IBM group’s director, infamously attributed their success to firing all the linguists. This talk looks at the history of speech recognition research as it was refashioned from a problem of simulating language to one of sorting data. Starting in the 1970s, speech recognition research shifted from efforts to study and simulate the processes of speech production and linguistic understanding to what researchers characterized as a “purely statistical” approach, organized around the technical and commercial demands of digital computing. I examine how the problem of automatic speech recognition, laden with the technical challenges and institutional legacies of acoustic engineering, helped bring language under the purview of data processing—and how, in the process, speech recognition research became critical in shaping the conceptual, economic, and technical terrain that gave rise to data-driven analytics and machine learning as privileged and pervasive forms of computational knowledge.