Search

A computational model for auditory word recognition is presented that enhances the model of Arnold et al. (2017). Real-valued features are extracted from the speech signal instead of discrete features. One-hot encoding for words’ meanings is replaced by real-valued semantic vectors, adding a small amount of noise to safeguard discriminability. Instead of learning with Rescorla-Wagner updating, we use multivariate multiple regression, which captures discrimination learning at the limit of experience. These new design features substantially improve prediction accuracy for words extracted from spontaneous conversations. They also provide enhanced temporal granularity, enabling the modeling of cohort-like effects. Clustering with t-SNE shows that the acoustic form space captures phone-like similarities and differences. Thus, wide learning with high-dimensional vectors and no hidden layers, and no abstract mediating phone-like representations is not only possible but achieves excellent performance that approximates the lower bound of human accuracy on the challenging task of isolated word recognition.

Classical linguistic, cognitive and engineering models for speech recognition and human auditory comprehension posit representations for sounds and words that mediate between the acoustic signal and interpretation. Recent advances in automatic speech recognition have shown, using deep learning, that state-of-the-art performance is obtained without such units. We present a cognitive model of auditory comprehension based on wide rather than deep learning that was trained on 20 to 80 hours of TV news broadcasts. Just as deep network models, our model is an end-to-end system that does not make use of phonemes and phonological wordform representations. Nevertheless, it performs well on the difficult task of single word identification (model accuracy 11.37%, Mozilla DeepSpeech: 4.45%). The architecture of the model is a simple two-layered wide neural network with weighted connections between the acoustic frequency band features as inputs and lexical outcomes (pointers to semantic vectors) as outputs. Model performance shows hardly any degredation when trained on speech in noise rather than on clean speech. Performance was further enhanced by adding a second network to a standard wide network. The present word recognition module is designed to become part of a larger system modeling the comprehension of running speech.