In this paper, we explore neural network models that learn to associate segments of spoken audio captions with the semantically relevant portions of natural images that they refer to. We demonstrate that these audio-visual associative localizations emerge from network-internal representations learned as a by-product of training to perform an image-audio retrieval task. Our models operate directly on the image pixels and speech waveform, and do not rely on any conventional supervision in the form of labels, segmentations, or alignments between the modalities during training. We perform analysis using the Places 205 and ADE20k datasets demonstrating that our models implicitly learn semantically coupled object and word detectors.
In this paper, we explore neural network models that learn to associate segments of spoken audio captions with the semantically relevant portions of natural images that they refer to. We demonstrate that these audio-visual associative localizations emerge from network-internal representations learned as a by-product of training to perform an image-audio retrieval task. Our models operate directly on the image pixels and speech waveform, and do not rely on any conventional supervision in the form of labels, segmentations, or alignments between the modalities during training. We perform analysis using the Places 205 and ADE20k datasets demonstrating that our models implicitly learn semantically-coupled object and word detectors.
Anuj Diwan, Layne Berry, Eunsol Choi, David Harwath, Kyle Mahowald. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. 2022.
Speech decoding-based brain-computer interfaces (Speech-BCIs) decode speech directly from brain signals, which have the potential to offer faster and natural communication to patients with locked-in syndrome than the current BCI-spellers. On account of the huge cognitive variance among subjects, most of the current speech-BCI models have focused on subject-dependent decoding where the training and evaluation of the decoding algorithms use data from the same participants. These models do not generalize across individuals and, thus, are limited by the small data size that can be obtained from a single participant. Few studies have attempted subject-independent decoding but the performances are sub-par at best and significantly lower than subject-dependent models. To address this issue, we evaluated imagined and overt speech decoding with magnetoencephalography (MEG) recordings of eight speakers in a generalizable subject-independent setting. We used recent domain adaptation techniques including feature augmentation and curriculum learning to introduce generalizability to the decoding model. Our results indicated that domain adaptation techniques can be efficient in subject-independent decoding. The best performance was obtained with a curriculum learning based adaptation technique that resulted in decoding accuracy was close to that in subject-dependent decoding. Our findings show the possibility of subject generalization in neural speech decoding.