Neural rhythms as priors of speech computations
Chandravadia, N.; Imam, N.
Show abstract
The transformation of speech into discrete linguistic representations forms the basis of speech recognition. Natural speech encodes cues at distinct timescales: phonetic features have modulation frequencies of 30-50 Hz, syllables and words around 4-7 Hz, and phrases 1-2 Hz. Strikingly, these frequencies mirror frequencies of endogenous network rhythms of the brain and synaptic time constants of the underlying neural circuits. Here, we suggest that endogenous brain rhythms serve as priors for speech recognition, encoding knowledge of speech structure in the dynamics of network computations. In a network of coupled oscillators, we find that speech is readily identified when characteristic frequencies of the oscillators match frequencies of circuit rhythms in the brain. When signal and circuit rhythms are mismatched, speech identification is impaired. Compared to a baseline recurrent neural network without intrinsic oscillations, the coupled oscillatory network has significantly higher performance in speech recognition across languages, but not in the recognition of signals that lack speech-like structure, such as urban sounds. Our results suggest a central computational role of brain rhythms in speech processing.
Matching journals
The top 4 journals account for 50% of the predicted probability mass.
Similar papers in this journal
- Two stages of bandwidth scaling drives efficient neural coding of natural sounds 96%
- Temporal resolution of spike coding in feedforward networks with signal convergence and divergence 95%
- Differential contributions of synaptic and intrinsic inhibitory currents to speech segmentation via flexible phase-locking in neural oscillators 95%
Similar papers in this journal
- Simple spectral transformations capture the contribution of peripheral processing to cortical responses to natural sounds 95%
- A solution to temporal credit assignment using cell-type-specific modulatory signals 94%
- Are place cells just memory cells? Memory compression leads to spatial tuning and history dependence 94%
"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.