Tag
This preprint introduces a generation-aligned diagnostic ladder that separates decision-rule misalignment from readout-coverage limitations in speech language models, showing that state decoding far exceeds generated accuracy in emotion recognition tasks.
This paper introduces MERaLiON-GR, a speech gender recognition model for English and Southeast Asian languages, fine-tuned from MERaLiON-SpeechEncoder-2 with LoRA and an ECAPA-TDNN head, achieving state-of-the-art performance across multilingual benchmarks.