NAPE
Summary
NAPE (Next-Audio-Patch-Embedding prediction) is a self-supervised audio representation-learning method that predicts each next log-mel spectrogram patch embedding from the causally visible patch sequence. Its reported pre-training recipe uses a causal Transformer, a predictor head, a one-position target shift, target stop-gradient, and negative cosine similarity without a reconstruction decoder, acoustic tokenizer, EMA teacher, or auxiliary regularization loss.
Evidence Status
NAPE was released as arXiv v1 on 2026-08-20 and was labelled under review on the first author’s publication page at the 2026-08-23 ingest. The official paper, project page, Apache-2.0 code, and three checkpoint endpoints are public, but the fresh release has no peer-reviewed venue, independent replication, tests, CI, or tagged version. The audited repository also has launcher/path documentation inconsistencies.
Main Result
Across Small/Base/Large encoders of approximately 19M/85M/303M parameters, NAPE’s reported fine-tuning scores improve monotonically on six audio/speech benchmarks. NAPE-L reaches 68.0% on IEMOCAP and ties the best listed AudioSet-2M mAP, while remaining below the best listed result on several other benchmark columns. The correct reading is competitive and task-specific state-of-the-art performance, not universal dominance.
Time-Series Relevance
NAPE provides cross-modal evidence that causal ordering over a 2D time-frequency grid is an actual modeling choice and that an independent patch-embedding target can be much easier to stabilize than a contextual encoder-layer target. It does not model multivariate numeric features, irregular event streams, actions, control inputs, interventions, or counterfactual futures.