MAGNeT by Meta

MAGNeT (Masked Audio Generation using a Single Non-Autoregressive Transformer) is a research model developed by Meta AI and collaborators for text-to-music and text-to-audio generation. It uses a single-stage, non-autoregressive transformer that operates directly over multiple streams of audio tokens. During training, the model predicts spans of masked tokens, and during inference it gradually constructs the output sequence using several decoding steps. MAGNeT introduces a novel rescoring method that leverages an external pre-trained model to rank predictions for later decoding steps, and a hybrid version that fuses autoregressive and non-autoregressive generation for the first few seconds. The model achieves comparable quality to existing baselines while being significantly faster (up to 7x faster than autoregressive approaches).