Experiments on context awareness and phone error propagation in human and machine speech recognition

A comparison of human speech recognition (HSR) and automatic speech recognition (ASR) is presented using a noisy continuous-speech corpus of null grammar or uniformly distributed unigram sentences, focusing on the differential tendency of machines vs. humans to propagate errors from an unclear phone to its neighbors. It is shown using controlled experiments that when given the same context for recognition in this case a vocabulary of a limited number of known words ASR makes as much as an order of magnitude more errors than HSR. The study provides evidence to contradict the claim made in recent literature that narrowing down the context of conversation and modeling of exceptional ordering of words is vital in achieving human-like accuracy by ASR. Using Chebyshev confidence intervals it is shown that ASR, but not HSR, propagates a phone recognition error from the phone to its neighbors at a rate significantly higher than chance.