Research

Walking Through Uncertainty: An Empirical Study of Uncertainty Estimation for Audio-Aware Large Language Models

arXiv 2026

note — Verification-based methods beat token-level likelihood baselines for uncertainty, which is neat. But it only tests Qwen Omni 2.5 and Flamingo — two fairly weak models — so I can't tell real generalization from overfitting to their reasoning quirks.

Accent Leakage in a Security Context

ICASSP 2026

Towards Language-Agnostic STIPA: Universal Phonetic Transcription to Support Language Documentation at Scale

EMNLP 2025

note — Whisper→IPA for endangered languages, and the confusion matrix stratified by place of articulation (nasals getting misread as diacritics!) is the gem. But why only Arabic? A 4-hour corpus is small enough to hand-annotate rather than curate by dictionary lookup, and they could have pretrained on adjacent languages.

VoiceBBQ: Investigating Effect of Content and Acoustics in Social Bias of Spoken Language Model

EMNLP 2025

note — Well-scoped as a short paper but shallow. I like the instinct that llama's frozen feature encoder blocks acoustic reasoning (making it better for textual bias probing) compared to qwen — but it only tests American and British accents and fairly underdeveloped LALMs.

Hearing the Order: Investigating Position Bias in Large Audio-Language Models

arXiv 2025

note — Good spread of LALM benchmarks plus permutation fixes for ordering bias. I wish it hypothesized why the bias exists, especially between near-identical models like Qwen vs Qwen-Omni. Still chewing on their Choice KL-Divergence metric.

Algorithmic Language Management

Nina Markl

Do Datasets Have Politics?

Scheuerman, Hanna, Denton

wav2vec-S: Adapting Pre-trained Speech Models for Streaming

Fu et al.

Phonetic Enhanced Language Modeling for TTS

Zhou et al.