Human listeners readily adjust to unfamiliar speakers and language varieties through exposure, but do these adaptation benefits extend to state-of-the-art spoken language models (SLMs)? We introduce a scalable framework that allows for in-context learning (ICL) in Phi-4 Multimodal (Phi-4-MM) using interleaved task prompts and audio-text pairs, and find that as few as 12 example utterances (~50 seconds) at inference time reduce word error rates by a relative 19.7% (1.2 pp.) on average across diverse English corpora. These improvements are most pronounced in low-resource varieties, when the context and target speaker match, and when more examples are provided—though scaling our procedure yields diminishing marginal returns to context length. Overall, we find that our novel ICL adaptation scheme (1) reveals a similar performance profile to human listeners, and (2) demonstrates consistent improvements to automatic speech recognition (ASR) robustness across diverse speakers and language backgrounds. While adaptation succeeds broadly, significant gaps remain for certain varieties, revealing where current models still fall short of human flexibility. We release our prompts and code on GitHub.

We introduce a novel fine-tuning approach that effectively primes transformer-based language models to detect rhetorical and psychological techniques within internet memes. Our end-to-end system retains multilingual and task-general capacities from pretraining stages while adapting to domain intricacies using an increasingly targeted set of examples– achieving competitive rankings across English, Bulgarian, and North Macedonian. We find that our monolingual post-training regimen is sufficient to improve task performance in 17 language varieties beyond equivalent zero-shot capabilities despite English-only data. To promote further research, we release our code publicly on GitHub.

We develop and probe a model for detecting the boundaries of prosodic chunks in untranscribed conversational English speech. The model is obtained by fine-tuning a Transformer-based speech-to-text (STT) model to integrate the identification of Intonation Unit (IU) boundaries with the STT task. The model shows robust performance, both on held-out data and on out-of-distribution data representing different dialects and transcription protocols. By evaluating the model on degraded speech data, and comparing it with alternatives, we establish that it relies heavily on lexico-syntactic information inferred from audio, and not solely on acoustic information typically understood to cue prosodic structure. We release our model as both a transcription tool and a baseline for further improvements in prosodic segmentation.

This study investigates the clustering of words into Part-of-Speech (POS) classes in Kolyma Yukaghir. In grammatical descriptions, lexical items are assigned to POS classes based on their morphological paradigms. Discursively, however, these classes share a fair amount of morphology. In this study, we turn to POS induction to evaluate if classes based on quantification of the distributions in which roots and affixes are used can be useful for language description purposes, and, if so, what those classes might be. We qualitatively compare clusters of roots and affixes based on four different definitions of their distributions. The results show that clustering is more reliable for words that typically bear more morphology. Additionally, the results suggest that the number of POS classes in Kolyma Yukaghir might be smaller than stated in current descriptions. This study thus demonstrates how unsupervised learning methods can provide insights for language description, particularly for highly inflectional languages.