Search for: All records

Experiments to understand the sensorimotor neural interactions in the human cortical speech system support the existence of a bidirectional flow of interactions between the auditory and motor regions. Their key function is to enable the brain to ‘learn’ how to control the vocal tract for speech production. This idea is the impetus for the recently proposed "MirrorNet", a constrained autoencoder architecture. In this paper, the MirrorNet is applied to learn, in an unsupervised manner, the controls of a specific audio synthesizer (DIVA) to produce melodies only from their auditory spectrograms. The results demonstrate how the MirrorNet discovers the synthesizer parameters to generate the melodies that closely resemble the original and those of unseen melodies, and even determine the best set parameters to approximate renditions of complex piano melodies generated by a different synthesizer. This generalizability of the MirrorNet illustrates its potential to discover from sensory data the controls of arbitrary motor-plants. 

Recent advancements in deep learning have led to drastic improvements in speech segregation models. Despite their success and growing applicability, few efforts have been made to analyze the underlying principles that these networks learn to perform segregation. Here we analyze the role of harmonicity on two state-of-the-art Deep Neural Networks (DNN)-based models- Conv-TasNet and DPT-Net [1],[2]. We evaluate their performance with mixtures of natural speech versus slightly manipulated inharmonic speech, where harmonics are slightly frequency jittered. We find that performance deteriorates significantly if one source is even slightly harmonically jittered, e.g., an imperceptible 3% harmonic jitter degrades performance of Conv-TasNet from 15.4 dB to 0.70 dB. Training the model on inharmonic speech does not remedy this sensitivity, instead resulting in worse performance on natural speech mixtures, making inharmonicity a powerful adversarial factor in DNN models. Furthermore, additional analyses reveal that DNN algorithms deviate markedly from biologically inspired algorithms [3] that rely primarily on timing cues and not harmonicity to segregate speech. 

Abstract Numerous studies have suggested that the perception of a target sound stream (or source) can only be segregated from a complex acoustic background mixture if the acoustic features underlying its perceptual attributes (e.g., pitch, location, and timbre) induce temporally modulated responses that are mutually correlated (or coherent), and that are uncorrelated (incoherent) from those of other sources in the mixture. This “temporal coherence” hypothesis asserts that attentive listening to one acoustic feature of a target enhances brain responses to that feature but would also concomitantly (1) induce mutually excitatory influences with other coherently responding neurons, thus enhancing (or binding) them all as they respond to the attended source; by contrast, (2) suppressive interactions are hypothesized to build up among neurons driven by temporally incoherent sound features, thus relatively reducing their activity. In this study, we report on EEG measurements in human subjects engaged in various sound segregation tasks that demonstrate rapid binding among the temporally coherent features of the attended source regardless of their identity (pure tone components, tone complexes, or noise), harmonic relationship, or frequency separation, thus confirming the key role temporal coherence plays in the analysis and organization of auditory scenes.