From 1939 to voice clones in 3 seconds — the full AI speech timeline and where it's heading

Dennis Klatt's voice became Stephen Hawking's voice. WaveNet scored 4.21 vs 3.86 for the best previous system. Tacotron 2 hit 4.53 vs 4.58 for real human speech. VALL-E cloned voices from 3 seconds. Microsoft refused to release VALL-E 2. Now Kokoro (82M params, trained for $400) competes with Eleven