Continuous Autoregressive Models with Noise Augmentation Avoid Error Accumulation

TL;DR

This paper presents a novel noise augmentation technique for Continuous Autoregressive Models (CAMs) that effectively mitigates error accumulation during long sequence generation, significantly improving audio quality in musical applications.

Contribution

The authors introduce a noise injection training method and a low-noise inference procedure to enhance the robustness of CAMs against error accumulation.

Findings

CAM outperforms existing methods in musical audio generation

Noise augmentation reduces error propagation over long sequences

Enhanced audio quality maintained in extended sequence generation

Abstract

Autoregressive models are typically applied to sequences of discrete tokens, but recent research indicates that generating sequences of continuous embeddings in an autoregressive manner is also feasible. However, such Continuous Autoregressive Models (CAMs) can suffer from a decline in generation quality over extended sequences due to error accumulation during inference. We introduce a novel method to address this issue by injecting random noise into the input embeddings during training. This procedure makes the model robust against varying error levels at inference. We further reduce error accumulation through an inference procedure that introduces low-level noise. Experiments on musical audio generation show that CAM substantially outperforms existing autoregressive and non-autoregressive approaches while preserving audio quality over extended sequences. This work paves the way for generating continuous embeddings in a purely autoregressive setting, opening new possibilities for real-time and interactive generative applications.