The Baum-Welch algorithm
Baum and co-authors described how to estimate the parameters of a hidden Markov process from observations, without ever seeing the states.
Why it matters
A model where the cause is hidden and only the effect is visible became trainable, and that defined speech recognition for thirty years.
The paper is purely mathematical and never mentions speech. But it gives the iterative procedure that raises likelihood at every step, later recognised as a special case of expectation maximisation. A speech signal fits the model well: articulation is hidden and only acoustics are observed. Without this result neither DRAGON nor the IBM systems would have had anything to train.