Abstract
General-purpose audio foundation models (FMs) have not been systematically evaluated for cross-water underwater acoustic target recognition (UATR), where transfer performance degrades catastrophically even for models trained on underwater acoustic data [2]. We benchmark eight public foundation models (wav2vec2-base/large, HuBERT-base, WavLM-base+, AST, CLAP-HTSAT, PANNs-CNN14, BEATs) as frozen feature extractors on the four UATR corpora of this series (258,893 clips), covering in-corpus evaluation and cross-corpus zero-target-label transfer: 12 directed corpus pairs × 8 models = 96 cells; probes use source-domain labels only. Findings: (1) in all 32 full-label in-corpus cells, frozen-FM probes beat the majority-class baseline in macro-F1 but mostly fall short of the trained ResNet-18 reference (BEATs slightly exceeds the trained reference on ShipsEar and leads foundation models on VTUAD in macro-F1, yet remains below the fine-tuned reference); (2) scaling up yields no consistent gain: wav2vec2-large never beats wav2vec2-base on both metrics on any corpus; (3) cross-corpus, the best frozen-FM probe beats zero-shot transfer of trained models in 11 of 12 directions, and WavLM's Oceanship→QiandaoEar22 probe (macro-F1 0.736) exceeds the supervised target-domain oracle (0.551); 5,000 permutation controls, however, place it at the 96.3rd–97.4th percentile of random readouts (Fig. 2); we therefore report it as a qualified observation, not a superiority claim; (4) a deployable source-validation selection rule beats the trained baseline in 7 of 12 directions; the five failures concentrate on VTUAD-sourced transfer, where source-validation rankings do not carry across waters. These are not iso-resource comparisons: frozen FM embeddings provide an adaptation-free, zero-target-label baseline, not a replacement for in-domain training.