What do end-to-end speech models learn about speaker, language and channel information? A layer-wise and neuron-level analysis

Deep neural networks are inherently opaque and challenging to interpret. Unlike hand-crafted feature-based models, we struggle to comprehend the concepts learned and how they interact within these models. This understanding is crucial not only for debugging purposes but also for ensuring fairness in ethical decision-making. In our study, we conduct a post-hoc functional interpretability analysis of pretrained speech models using the probing framework (Hupkes et al., 2018). Specifically, we analyze utterance-level representations of speech models trained for various tasks such as speaker recognition and dialect identification. We conduct layer and neuron-wise analyses, probing for speaker, language, and channel properties. Our study aims to answer the following questions: (i) what information is captured within the representations? (ii) how is it represented and distributed? and (iii) can we identify a minimal subset of the network that possesses this information? Our results reveal several novel findings, including: (i) channel and gender information are distributed across the network, (ii) the information is redundantly available in neurons with respect to a task, (iii) complex properties such as dialectal information are encoded only in the task-oriented pretrained network, (iv) and is localized in the upper layers, (v) we can extract a minimal subset of neurons encoding the pre-defined property, (vi) salient neurons are sometimes shared between properties, (vii) our analysis highlights the presence of biases (for example gender) in the network. Our cross-architectural comparison indicates that: (i) the pretrained models capture speaker-invariant information, and (ii) CNN models are competitive with Transformer models in encoding various understudied properties.