Frederick Jelinek (1932 - September 14, 2010) was a Czech-American researcher whose work moved speech recognition and machine translation away from hand-built linguistic rules and toward statistics. According to the obituary by Mark Liberman in the journal Computational Linguistics, Jelinek received his information-theory PhD from MIT in 1962, taught at Cornell until 1972, led speech recognition research at IBM from 1972 to 1993, and then directed the Center for Language and Speech Processing at Johns Hopkins until his death.
Liberman frames Jelinek’s contribution as “persuading the fields of speech and language engineering to adopt statistical methods and the noisy channel model, returning to the path opened by Claude Shannon in 1948.” When Jelinek arrived at IBM, information theory had fallen out of favor for language tasks - undercut by Chomsky’s argument that statistics could not capture syntax and by John Pierce’s influential attacks on speech-recognition funding. Jelinek’s group rebuilt the case empirically, showing that probabilistic models trained on data outperformed rule-based systems.
Just as lasting was the research paradigm he championed: the competitive evaluation of alternative algorithms against shared training and test sets using fixed, automatically computed metrics. This “common task” method, adopted by DARPA in the mid-1980s, became the standard way the entire field measures progress, and IBM under Jelinek donated its aligned Canadian Hansards parallel corpus to seed shared data resources.
Jelinek is most often quoted for the line “Every time I fire a linguist, the performance of the speech recognizer goes up.” The exact wording and occasion are famously uncertain - Jelinek himself recalled it differently on different occasions and could not pin down when he first said it - so the quote is best treated as a paraphrase of his stance rather than a verified utterance. The sentiment, that data-driven methods beat linguistic intuition, captures the revolution he led.