A new research paper from Hugging Face introduces NAPE (Next-Audio-Patch-Embedding prediction), a self-supervised learning framework for audio that relies on a simple causal paradigm. Unlike many existing methods that use complex pre-training recipes, NAPE trains a causal Transformer to predict each next patch embedding of a log-mel spectrogram from the previous ones, using only causal masking and stop-gradient as training signals.
The design is intentionally minimalist, avoiding reconstruction decoders, acoustic tokenizers, student-teacher setups, and auxiliary regularization losses. This approach aligns with the autoregressive prediction philosophy that has driven progress in language modeling and visual representation learning, suggesting that a unified pre-training interface can transfer across modalities.
Across six audio and speech benchmarks, NAPE achieves state-of-the-art fine-tuning performance on several tasks, scales consistently across encoder sizes, and yields strong linear-probing results. The authors also note that NAPE produces structured attention patterns without explicit supervision, indicating that the model learns meaningful representations of audio structure.
The paper is available on Hugging Face and has been recommended alongside related works such as AV-JEPA, DINO-A, and MJEPA, which explore similar predictive architectures for audio and audio-visual learning.