Fig 1. A generative model for on-line syllable recognition with rhythmic state-dependent precisions.

The diagram in subpanel a) shows the simplified schematics and hierarchical message passing of Precoss-β. The lower panel shows the input to the model. As in the original model (Precoss), the input consists of the speech slow amplitude modulation (on the left) and the auditory spectrogram (on the right). At the top level, the theta module tracks the slow amplitude modulation in the input and feeds it to a theta oscillator. The instantaneous frequency of the theta oscillator and the Gaussian pulses associated with the predefined phases signal speech rate and syllable onset information to the gamma units. Together, the gamma and syllable units produce the auditory spectrogram in the input, based on the spectrotemporal patterns stored in the model’s memory (subpanel c). The gamma units make temporal predictions about the spectrotemporal patterns of syllables (as many as there are syllables in the input sentence), while the syllable units accumulate evidence about each syllable in the sentence. Depending on the phase of the oscillating PEP, the model changes the precision of the syllable and gamma units, modulating the influence of the corresponding prediction errors on the dynamics of the hidden states. Depending on the phase of the precision units (highlighted by the grey rectangle), either syllable or gamma units get higher precision. The arrows represent a message passing between levels of the model hierarchy (top-down predictions in grey and bottom-up prediction errors in magenta) and lateral, within-level connections (green arrows). Subpanel b) shows an example of the dynamics of syllable and gamma units from the model simulations. The top panel shows the accumulated evidence for different syllables (color coded) in the input sentence (colored bars on top of syllable units represent syllables in the input sentence), while the bottom panel shows the sequential activation of gamma units. Subpanel d) shows the simplified diagram of the model, where θ, γ and ω represent theta module, gamma and syllable units respectively. The π represents the oscillating precision (arrows indicate the units whose precision is controlled).