Zero-shot TTS just crossed a line many thought was still years away: Microsoft's VALL-E 2 is the first neural codec language model to hit human parity on both LibriSpeech and VCTK across robustness, naturalness, and speaker similarity. This detailed review unpacks how the team got there, focusing on two deceptively simple architectural tweaks that quietly fix the biggest failure modes of the original VALL-E — the runaway decoding loops and the brittle output on long, complex, or repetitive input sentences that used to break the model.

The core moves are Repetition Aware Sampling, which reshapes nucleus sampling by conditioning on token repetition history to stabilize decoding and kill infinite loops, and Grouped Code Modeling, which chunks codec tokens into groups to shrink sequence length, accelerate inference, and address the challenges of long sequence modeling. The review walks through why these matter for anyone building production TTS: consistently higher-quality output even on traditionally hard sentences, stronger speaker similarity, and a plausible path to accessibility applications like restoring voices for people with ALS or aphasia. If you work on codec LMs, neural vocoders, or zero-shot voice cloning, this breakdown is where to start before diving into the paper.