Rounding and the third formant: An assumption revisited with deep neural networks

A student of the phonetic sciences who consults a relevant coursebook will find there the claim that F3 is related to the degree of lip rounding. However, the exact relationship between the two is variable and inconsistent. Here, we present results of an exploratory computational-acoustic model where changes to F3 are interpreted in relationship to changes to vocal tract area transfer function of a large sample (N = 10,000) of systematically varied tube model sequences. The results show a consistent pattern: constriction at the front of the vocal tract, followed by an open area and a secondary constriction, systematically shapes F3. This pattern aligns with acoustic signatures of both rounding and rhoticity, without presupposing phonological categories. These results offer a principled acoustic account of why the traditional link between rounding and lowered F3 varies in natural speech. Future work will explore whether certain vocal tract regions are more acoustically “sensitive” than others.