Our research consisted of two studies focusing on the probability of humans being able to perceive the difference between valence of human vocalizations of high (pain, pleasure and fear) and low intensity (laugh and neutral speech). The first study was conducted online and used a large sample (n=902) of respondents. The second study was conducted in a laboratory setting and involved a stress induction procedure. For both, the task was to categorize whether the human vocalization was rated positive, neutral or negative. Stimuli were audio records extracted from freely downloadable online videos and can be considered semi-naturalistic. Each rating participant (rater) was presented with five audio records (stimuli) of five females and of five males. All raters were presented with the stimuli twice (so as to statistically estimate the consistency of the ratings). Using a Bayesian statistical approach, we could test for consistencies and due-to-chance probabilities. The outcomes support the prediction that the results (ratings) are repeatable (not due to chance) but incorrectly attributed, decreasing the communication value of the expressions of fear, pain, and pleasure. Stress induction (in study two conducted on 28 participants) did have an impact on the ratings of male neutral and laugh – it caused decrease in correct attribution.