Training-Free Spectral Fingerprints of Voice Processing in Transformers