"""Context-specialized speaker prediction: 320 ms / 2 s / 3.04 s."""