Given the fast development of analysis techniques for NLP and speech processing systems, few systematic studies have been conducted to compare the strengths and weaknesses of each method. As a step in this direction we study the case of representations of phonology in neural network models of spoken language. We use two commonly applied analytical techniques, diagnostic classifiers and representational similarity analysis, to quantify to what extent neural activation patterns encode phonemes and phoneme sequences. We manipulate two factors that can affect the outcome of analysis. First, we investigate the role of learning by comparing neural activations extracted from trained versus randomly-initialized models. Second, we examine the temporal scope of the activations by probing both local activations corresponding to a few milliseconds of the speech signal, and global activations pooled over the whole utterance. We conclude that reporting analysis results with randomly initialized models is crucial, and that global-scope methods tend to yield more consistent results and we recommend their use as a complement to local-scope diagnostic methods.

We propose a novel method for computationally efficient audio-visual voice activity detection (VAD) where visual temporal information is provided by an energy efficient event-camera (EC). Unlike conventional cameras, ECs perform on-chip low-power pixel-level change detection, adapting the sampling frequency to the dynamics of the activity in the visual scene and removing redundancy, hence enabling energy and computational efficiency. In our VAD pipeline, first, lip activity is located and detected jointly by a probabilistic estimation after spatio-temporal filtering. Then, over the lips, a feather-weight speech-related lip motion detection is performed with minimum false negative rate to activate a highly accurate but expensive acoustic deep neural networks-based VAD. Our experiments show that ECs are accurate at detecting and locating lip activity; and EC-driven VAD can result in considerable savings in computations as well as can substantially reduce false positive rates in low acoustic signal-to-noise ratio conditions.