Search for: All records

Although the application of deep learning to automatic speech recognition (ASR) has resulted in dramatic reductions in word error rate for languages with abundant training data, ASR for languages with few resources has yet to benefit from deep learning to the same extent. In this paper, we investigate various methods of acoustic modeling and data augmentation with the goal of improving the accuracy of a deep learning ASR framework for a low-resource language with a high baseline word error rate. We compare several methods of generating synthetic acoustic training data via voice transformation and signal distortion, and we explore several strategies for integrating this data into the acoustic training pipeline. We evaluate our methods on an indigenous language of North America with minimal training resources. We show that training initially via transfer learning from an existing high-resource language acoustic model, refining weights using a heavily concentrated synthetic dataset, and finally fine-tuning to the target language using limited synthetic data reduces WER by 15% over just transfer learning using deep recurrent methods. Further, we show improvements over traditional frameworks by 19% using a similar multistage training with deep convolutional approaches. 

Documenting endangered languages supports the historical preservation of diverse cultures. Automatic speech recognition (ASR), while potentially very useful for this task, has been underutilized for language documentation due to the challenges inherent in building robust models from extremely limited audio and text training resources. In this paper, we explore the utility of supplementing existing training resources using synthetic data, with a focus on Seneca, a morphologically complex endangered language of North America. We use transfer learning to train acoustic models using both the small amount of available acoustic training data and artificially distorted copies of that data. We then supplement the language model training data with verb forms generated by rule and sentences produced by an LSTM trained on the available text data. The addition of synthetic data yields reductions in word error rate, demonstrating the promise of data augmentation for this task. 

Humans routinely extract important information from images and videos, relying on their gaze. In contrast, computational systems still have difficulty annotating important visual information in a human-like manner, in part because human gaze is often not included in the modeling process. Human input is also particularly relevant for processing and interpreting affective visual information. To address this challenge, we captured human gaze, spoken language, and facial expressions simultaneously in an experiment with visual stimuli characterized by subjective and affective content. Observers described the content of complex emotional images and videos depicting positive and negative scenarios and also their feelings about the imagery being viewed. We explore patterns of these modalities, for example by comparing the affective nature of participant-elicited linguistic tokens with image valence. Additionally, we expand a framework for generating automatic alignments between the gaze and spoken language modalities for visual annotation of images. Multimodal alignment is challenging due to their varying temporal offset. We explore alignment robustness when images have affective content and whether image valence influences alignment results. We also study if word frequency-based filtering impacts results, with both the unfiltered and filtered scenarios performing better than baseline comparisons, and with filtering resulting in a substantial decrease in alignment error rate. We provide visualizations of the resulting annotations from multimodal alignment. This work has implications for areas such as image understanding, media accessibility, and multimodal data fusion.