A study was conducted and published at the University of Wisconsin-Madison in 2013, Inherent Emotional Quality of Human Speech Sounds, in which it was demonstrated...
...that certain strings of English phonemes have an inherent, non-arbitrary emotional valence which can be predicted on the basis of dynamic changes in acoustic features.
It was partly based on the idea that animal vocalizations with higher formant positions were considered to be smaller organisms and hence less threatening than animals with vocalization with lower formant positions.
To remove any "biasing", nonsense words were used with both upward and downward F1-F2 shifts, and subjects were asked to associate the "words" with pictures, much like in the Kolher study (1929). Examples of the "words" are shown below.
And this is an example of "words" + pictures:
The conclusions were that:
The data presented here effectively outline a formula for constructing words and non-words that implicitly conjure positive or negative emotion. Accordingly we see potential applications of this study to a variety of educational, social, clinical, and marketing contexts.
I do not know if this study has ever been replicated; I will leave that up to you to follow up.
Below is the link to the PDF of the complete Meyers-Shulz study.