This paper presents a framework for systematic probing of discrete speech token representations in spoken language models (SLMs). We propose three complementary components: a distributional divergence analysis testing whether an attribute is reflected in token usage, token-based classifiers to quantify recoverability and an attribute-conditioned representation analysis revealing phonetic attribute realizations. As a demonstration we apply these probes to tokenizer outputs and model generations from CosyVoice2 and SparkTTS on LibriTTS-R and VCTK. We find that gender is encoded in their respective tokens but in different forms - the signal is more stable across stages and datasets in CosyVoice2, whereas SparkTTS shows weaker cross-stage consistency and stronger pause/prosody-related effects. Exploratory probes of valence, arousal, and dominance are weaker and less consistent. These results show that discrete speech tokens retain speaker-related information in different ways across architectures and that the proposed framework provides an interpretable basis for comparing token representations across spoken language modeling pipelines.

This paper presents a systematic evaluation of nearest neighbors across semantic representation spaces in both textual and visual modalities. We focus on nominal concepts with varying concreteness levels, and apply a neighborhood overlap measure to compare these target concepts differing in their linguistic and perceptual nature. We find that alignment is primarily determined by modality, and additionally by level of concreteness: Models from the same modality show stronger alignment than cross-modal models, and spaces of concrete concepts show stronger alignment than those of abstract ones. Overall, larger neighborhood size strengthens the alignment between spaces.