This tries to prioritize actual voice to decide the waveform, and clamps noise to zero to ensure the waveform doesn't have a perceptually noisy base. In theory this better matches the overall voice message content.