What is VoiceNet? VoiceNet is a set of models that listen to a short speech clip and estimate 57 perceptual voice & speech dimensions — human-interpretable qualities such as voice age, arousal, warmth, brightness, tempo, breathiness, speaking style, resonance placement and more. Each dimension is described on an ordinal scale from 0 to 6 (a few use shorter scales), where the numbers correspond to plain-language rubric levels (for example, tempo 0 = glacially slow … 6 = auctioneer-fast). Every prediction below comes from a single audio embedding per clip, produced by the VoiceCLAP-commercial encoder, which small model "heads" then read to score each dimension. Two extra models are shown alongside: a Genuineness score (0–6, how authentic vs. performed the delivery sounds) and a Vocal-Burst Blend score (0–10, how naturally a laugh / sob / gasp / sigh blends into the surrounding speech).
Regression vs. classification. For every dimension we show two predictions. Regression gives a smooth estimate — a continuous number (e.g. 3.4) that can sit between the rubric levels, good for seeing subtle differences. Classification instead commits to a single hard bucket (a whole number 0–6) and comes with the exact plain-language description of that level. Regression is like a dimmer switch; classification is like picking the closest labelled notch.