Experiments in speech recognition over the telephone network

Moving a speech recognition algorithm out of the laboratory and into a real-world setting introduces many variables that have the potential to degrade performance significantly. Specifically, algorithms that explicitly or implicitly measure the difference between the power spectra of frames of an unknown utterance and a reference template will be negatively affected by sources of spectral distortion unrelated to linguistic distinctions. We have performed a series of experiments aimed at a deeper understanding of several of the variables that come into play when implementing one of the best-performing algorithms, that of Itakura [1], in a particularly harsh real-world environment, that of the telephone network.