Turning audio embeddings into speech recognition: a frozen EmbeddingGemma 2 audio encoder plus a 9.66M-parameter GRU-CTC reader, no encoder fine-tuning
-
Updated
Oct 16, 2026 - Python
Turning audio embeddings into speech recognition: a frozen EmbeddingGemma 2 audio encoder plus a 9.66M-parameter GRU-CTC reader, no encoder fine-tuning
World model on a Boolean hypercube: frozen reservoir encoders and a locally connected predictor that learns how the latent code moves under an action. C++20, no dependencies.
Do frozen image encoders transfer to radar images outside glaciers? Part two of a SAR transfer series. Sentinel-1 glacial lakes, snow, forest types and alpine glaciers with frozen satellite DINOv3 and C-RADIOv4, a linear probe and a small decoder, and frozen-feature views.
Bengali coreference resolution under low-resource constraints: data quality vs. quantity, augmentation, cross-lingual transfer, and embedding collapse in multilingual encoders. Code for our KONVENS 2026 paper (arXiv:2610.35114).
Can a distilled student match its frozen teachers on radar? Part four of a SAR transfer series. A C-RADIO-style ViT-B student learns from SigLIP2, satellite DINOv3, PE-Spatial and TerraFM-B reading Sentinel-1, scored with frozen features on glacial lakes, snow, forest types and alpine glaciers.
Do frozen image encoders transfer to SAR? Glacier calving fronts on CaFFe with frozen DINOv3 (satellite and photo), C-RADIO, TerraMind and SARATR-X and small heads, against a U-Net from scratch; pixel size vs radar band; frozen-feature views.
Frozen encoder + Mahalanobis prototype for class-incremental intent classification. 50+ experiments across BANKING77, CLINC150, HWU64, AG News. Matches fine-tuned baselines at 5MB state with zero forgetting, order-invariance, 455 QPS.
To associate your repository with the frozen-encoder topic, visit your repo's landing page and select "manage topics."