skip to main content

Title: Exploration of Acoustic and Lexical Cues for the INTERSPEECH 2020 Computational Paralinguistic Challenge
In this paper, we investigate various acoustic features and lexical features for the INTERSPEECH 2020 Computational Paralinguistic Challenge. For the acoustic analysis, we show that the proposed FV-MFCC feature is very promising, which has very strong prediction power on its own, and can also provide complementary information when fused with other acoustic features. For the lexical representation, we find that the corpus-dependent TF.IDF feature is by far the best representation. We also explore several model fusion techniques to combine different modalities together, and propose novel SVM models to aggregate the chunk-level predictions to the narrative-level predictions based on the chunk-level decision functionals. Finally we discuss the potential for improving prediction by combining the lexical and acoustic modalities together, and we find that fusion of lexical and acoustic modalities do not lead to consistent improvements over elderly Arousal, but substantially improve over the Valence. Our methods significantly outperform the official baselines on the test set in the participated Mask and Elderly Sub-challenges. We obtain an UAR of 75.1%, 54.3%, and 59.0% on the Mask, Elderly Arousal and Valence prediction tasks respectively.
; ; ; ;
Award ID(s):
Publication Date:
Journal Name:
Page Range or eLocation-ID:
2092 to 2096
Sponsoring Org:
National Science Foundation
More Like this
  1. In this paper, we propose a deep multimodal fusion network to fuse multiple modalities (face, iris, and fingerprint) for person identification. The proposed deep multimodal fusion algorithm consists of multiple streams of modality-specific Convolutional Neural Networks (CNNs), which are jointly optimized at multiple feature abstraction levels. Multiple features are extracted at several different convolutional layers from each modality-specific CNN for joint feature fusion, optimization, and classification. Features extracted at different convolutional layers of a modality-specific CNN represent the input at several different levels of abstract representations. We demonstrate that an efficient multimodal classification can be accomplished with a significant reduction in the number of network parameters by exploiting these multi-level abstract representations extracted from all the modality-specific CNNs. We demonstrate an increase in multimodal person identification performance by utilizing the proposed multi-level feature abstract representations in our multimodal fusion, rather than using only the features from the last layer of each modality-specific CNNs. We show that our deep multi-modal CNNs with multimodal fusion at several different feature level abstraction can significantly outperform the unimodal representation accuracy. We also demonstrate that the joint optimization of all the modality-specific CNNs excels the score and decision level fusions of independently optimized CNNs.
  2. Regularization plays a key role in improving the prediction of emotions using attributes such as arousal, valence and dominance. Regularization is particularly important with deep neural networks (DNNs), which have millions of parameters. While previous studies have reported competitive performance for arousal and dominance, the prediction results for valence using acoustic features are significantly lower. We hypothesize that higher regularization can lead to better results for valence. This study focuses on exploring the role of dropout as a form of regularization for valence, suggesting the need for higher regularization. We analyze the performance of regression models for valence, arousal and dominance as a function of the dropout probability. We observe that the optimum dropout rates are consistent for arousal and dominance. However, the optimum dropout rate for valence is higher. To understand the need for higher regularization for valence, we perform an empirical analysis to explore the nature of emotional cues conveyed in speech. We compare regression models with speakerdependent and speaker-independent partitions for training and testing. The experimental evaluation suggests stronger speaker dependent traits for valence. We conclude that higher regularization is needed for valence to force the network to learn global patterns that generalize across speakers.
  3. This study proposes a data fusion and deep learning (DL) framework that learns high-level traffic features from network-level images to predict large-scale, multi-route, speed and volume of connected vehicles (CVs). We present a scalable and parallel method of processing statewide CVs’ trajectory data that leads to real-time insights on the micro-scale in time and space (two-dimensional (2D) arrays) on graphics processing unit (GPUs) using the Nvidia rapids framework and dask parallel cluster, which provided a 50× speed-up in the data extraction, transform and load (ETL). A UNet model is then applied to perform feature extraction and multi-route speed and volume channels over a multi-step prediction horizon. The accuracy and robustness of the proposed model are evaluated by taking different road types, times of day and image snippets and comparing the model to benchmarks: Convolutional Long–Short-Term Memory (ConvLSTM) and a historical average (HA). The results show that the proposed model outperforms benchmarks with an average improvement of 15% over ConvLSTM and 65% over the HA. Comparing the image snippets from each prediction model to the actual image shows that image textures were highly similar in UNet to the benchmark models used. UNet’s dominance in performing image predictions was also evident inmore »multi-step forecasting, where the increase in errors was relatively minimal over longer prediction horizons.« less
  4. Understanding neural function often requires multiple modalities of data, including electrophysiogical data, imaging techniques, and demographic surveys. In this paper, we introduce a novel neurophysiological model to tackle major challenges in modeling multimodal data. First, we avoid non-alignment issues between raw signals and extracted, frequency-domain features by addressing the issue of variable sampling rates. Second, we encode modalities through “cross-attention” with other modalities. Lastly, we utilize properties of our parent transformer architecture to model long-range dependencies between segments across modalities and assess intermediary weights to better understand how source signals affect prediction. We apply our Multimodal Neurophysiological Transformer (MNT) to predict valence and arousal in an existing open-source dataset. Experiments on non-aligned multimodal time-series show that our model performs similarly and, in some cases, outperforms existing methods in classification tasks. In addition, qualitative analysis suggests that MNT is able to model neural influences on autonomic activity in predicting arousal. Our architecture has the potential to be fine-tuned to a variety of downstream tasks, including for BCI systems.
  5. We propose a method for supervised learning with multiple sets of features (“views”). The multiview problem is especially important in biology and medicine, where “-omics” data, such as genomics, proteomics, and radiomics, are measured on a common set of samples. “Cooperative learning” combines the usual squared-error loss of predictions with an “agreement” penalty to encourage the predictions from different data views to agree. By varying the weight of the agreement penalty, we get a continuum of solutions that include the well-known early and late fusion approaches. Cooperative learning chooses the degree of agreement (or fusion) in an adaptive manner, using a validation set or cross-validation to estimate test set prediction error. One version of our fitting procedure is modular, where one can choose different fitting mechanisms (e.g., lasso, random forests, boosting, or neural networks) appropriate for different data views. In the setting of cooperative regularized linear regression, the method combines the lasso penalty with the agreement penalty, yielding feature sparsity. The method can be especially powerful when the different data views share some underlying relationship in their signals that can be exploited to boost the signals. We show that cooperative learning achieves higher predictive accuracy on simulated data and realmore »multiomics examples of labor-onset prediction. By leveraging aligned signals and allowing flexible fitting mechanisms for different modalities, cooperative learning offers a powerful approach to multiomics data fusion.« less