Microsoft's neural codec language model, VALL-E, tokenises speech and uses algorithms to create waveforms that sound like the speaker while keeping their timbre and emotional tone.