The Time-Invariant String Kernel (TISK) model of spokenword recognition (Hanngan et al., 2013) is an interactiveactivation model like TRACE (McClelland & Elman, 1986).However, it uses orders of magnitude fewer nodes andconnections because it replaces TRACE's time-specificduplicates of phoneme and word nodes with time-invariantnodes based on a string kernel representation (essentially aphoneme-by-phoneme matrix, where a word is encoded as byall ordered open diphones it contains; e.g., cat has /kæ/, /æt/,and /kt/). Hannagan et al. (2013) showed that TISK behavessimilarly to TRACE in the time course of phonologicalcompetition and even word-specific recognition times.However, the original implementation did not includefeedback from words to diphone nodes, precluding simulationof top-down effects. Here, we demonstrate that TISK can beeasily adapted to lexical feedback, affording simulation oftop-down effects as well as allowing the model todemonstrate graceful degradation given noisy inputs.

Pervasive behavioral and neural evidence for predictiveprocessing has led to claims that language processing dependsupon predictive coding. In some cases, this may reflect aconflation of terms, but predictive coding formally is acomputational mechanism where only deviations from top-down expectations are passed between levels of representation.We evaluate three models’ ability to simulate predictiveprocessing and ask whether they exhibit the putative hallmarkof formal predictive coding (reduced signal when inputmatches expectations). Of crucial interest, TRACE, aninteractive activation model that does not explicitly implementprediction, exhibits both predictive processing and model-internal signal reduction. This may indicate that interactiveactivation is functionally equivalent or approximant topredictive coding, or that caution is warranted in interpretingneural signal reduction as diagnostic of predictive coding.

Despite the lack of invariance problem (the many-to-manymapping between acoustics and percepts), we experiencephonetic constancy and typically perceive what a speakerintends. Models of human speech recognition have side-stepped this problem, working with abstract, idealized inputsand deferring the challenge of working with real speech. Incontrast, automatic speech recognition powered by deeplearning networks have allowed robust, real-world speechrecognition. However, the complexities of deep learningarchitectures and training regimens make it difficult to usethem to provide direct insights into mechanisms that maysupport human speech recognition. We developed a simplenetwork that borrows one element from automatic speechrecognition (long short-term memory nodes, which providedynamic memory for short and long spans). This allows thenetwork to learn to map real speech from multiple talkers tosemantic targets with high accuracy. Internal representationsemerge that resemble phonetically-organized responses inhuman superior temporal gyrus, suggesting that the modeldevelops a distributed phonological code despite no explicittraining on phonetic or phonemic targets. The ability to workwith real speech is a major advance for cognitive models ofhuman speech recognition.