ESTIMATING CHANGES IN SPEECH METRICS INDICATIVE OF FATIGUE LEVELS

摘要

In this paper, we are presenting a novel approach to estimate fatigue levels of train conductors, by analyzing the speech signal. An independent neural network joined with a Markov Model, will output the probability density, which illustrates the likelihood of the result of the first step to be accurate. Vigilance research has shown that, for most operators engaged in attention-intensive and monotonous tasks, retaining a constant level of alertness is almost impossible. Sleeping disorders, reduced hours of rest and disrupted circadian rhythms amplify this effect and lead to significantly increased fatigue levels. Increased fatigue levels manifest themselves in alterations of speech metrics, as compared to alert states of mind. To make a decision about the level of fatigue, we are proposing an alertness estimation system which uses speech metrics to generate a fatigue quotient indicative of the fatigue level. A speech pre-processor extracts metrics such as speech duration, word production rate and speech intensity from a continuous speech signal and uses a Fuzzy Logic algorithm to generate the fatigue quotient at any moment in time when speech is present. However, the nature of human interaction introduces levels of uncertainty, which make fatigue level recognition difficult. In other words, even with a perfectly trained neural network and Fuzzy Logic algorithm, we cannot make definite conclusions about the level of alertness. The reason being, that there is no guarantee that the estimated level of alertness is robust for a certain amount of time and didn't come from drinking half a cup of coffee. Moreover, coming up with a perfect model of speech-fatigue (i.e. input-output) for humans, to train the Fuzzy algorithm is almost impossible. For this reason the study of "Risk and Uncertainty" is an integral part of this research. Motivated by the distinction between "risk" (randomness that can be fully captured by probability and statistics) and "uncertainty" (all other types of randomness), we propose a fine taxonomy: fully reducible, partially reducible, and irreducible uncertainty, that can explain some of the key differences between long term alertness and a short term change of state that makes the operator alert. An experimental study is conducted where a hyper articulated speech signal with three different levels of simulated fatigue is analyzed by the algorithm and a probability density function is assigned to the fatigue quotient to take the risk and uncertainty into account and make the overall result more reliable.