“Mati Staniszewski predicts that fused or speech-to-speech models will likely grow to tens or hundreds of billions of parameters as they combine language model and voice generation capabilities.”