We present the CATENG systems submitted to the IWSLT 2026 Dialectal and Low-Resource Speech Translation shared task for the Catalan–English (CA–EN) pair. Although Catalan is not strictly low-resource, its dialectal diversity and relative under-representation in speech technology make it a challenging setting. We evaluate three unconstrained systems: two cascaded approaches combining ASR and MT, and one end-to-end model. Our primary system uses a Mamba-based ASR (ConMamba) with a fine-tuned NLLB-200 MT model, while a contrastive system replaces the ASR with Whisper-v3; we also evaluate an end-to-end SpeechT5 model with data augmentation. Experiments are conducted on the IWSLT 2026 Catalan dataset (15 hours), complemented with large-scale parallel text. Results show that cascaded systems outperform end-to-end ST, with Whisper-v3 + NLLB achieving 44.7 BLEU and 65.1 chrF. We find that performance is primarily constrained by ASR quality rather than MT capacity, and that Mamba-based ASR models provide competitive results, highlighting the importance of robust speech representations and dialectal coverage for Catalan–English speech translation.

Speech information in a pretrained wav2vec2.0 model is usually leveraged through its encoder, which has at least 95M parameters, being not so suitable for small footprint Keyword Spotting. In this work, we show an efficient way of profiting from wav2vec2.0’s linguistic knowledge, by recycling the phonetic information encoded in its latent codebook, which has been typically thrown away after pretraining. We do so by transferring the codebook as weights for the latent bottleneck of a Keyword Spotting Perceiver, thus initializing such model with phonetic embeddings already. The Perceiver design relies on cross-attention between these embeddings and input data to generate better representations. Our method delivers accuracy gains compared to random initialization, at no latency costs. Plus, we show that the phonetic embeddings can easily be downsampled with k-means clustering, speeding up inference in 3.5 times at only slight accuracy penalties.