This work presents a novel approach for speaker diarization to leverage lexical information provided by automatic speech recognition. We propose a speaker diarization system that can incorporate word-level speaker turn probabilities with speaker embeddings into a speaker clustering process to improve the overall diarization accuracy. To integrate lexical and acoustic information in a comprehensive way during clustering, we introduce an adjacency matrix integration for spectral clustering. Since words and word boundary information for word-level speaker turn probability estimation are provided by a speech recognition system, our proposed method works without any human intervention for manual transcriptions. We show that the proposed method improves diarization performance on various evaluation datasets compared to the baseline diarization system using acoustic information only in speaker embeddings.
Many text classification tasks are domain-dependent, and various domain adaptation approaches have been proposed to predict unlabeled data in a new domain. Domain-adversarial neural networks (DANN) and their variants have been used widely recently and have achieved promising results for this problem. However, most of these approaches assume that the label proportions of the source and target domains are similar, which rarely holds in most real-world scenarios. Sometimes the label shift can be large and the DANN fails to learn domain-invariant features. In this study, we focus on unsupervised domain adaptation of text classification with label shift and introduce a domain adversarial network with label proportions estimation (DAN-LPE) framework. The DAN-LPE simultaneously trains a domain adversarial net and processes label proportions estimation by the confusion of the source domain and the predictions of the target domain. Experiments show the DAN-LPE achieves a good estimate of the target label distributions and reduces the label shift to improve the classification performance.
An objective understanding of media depictions, such as about inclusive portrayals of how much someone is heard and seen on screen in film and television, requires the machines to discern automatically who, when, how and where someone is talking. Media content is rich in multiple modalities such as visuals and audio which can be used to learn speaker activity in videos. In this work, we present visual representations that have implicit information about when someone is talking and where. We propose a crossmodal neural network for audio speech event detection using the visual frames. We use the learned representations for two downstream tasks: i) audio-visual voice activity detection ii) active speaker localization in video frames. We present a state-of-the-art audio-visual voice activity detection system and demonstrate that the learned embeddings can effectively localize to active speakers in the visual frames.
In this study, we propose a new spectral clustering framework that can auto-tune the parameters of the clustering algorithm in the context of speaker diarization. The proposed framework uses normalized maximum eigengap (NME) values to estimate the number of clusters and the parameters for the threshold of the elements of each row in an affinity matrix during spectral clustering, without the use of parameter tuning on the development set. Even through this hands-off approach, we achieve a comparable or better performance across various evaluation sets than the results found using traditional clustering methods that apply careful parameter tuning and development data. A relative improvement of 17% in the speaker error rate on the well-known CALLHOME evaluation set shows the effectiveness of our proposed spectral clustering with auto-tuning.
Machine learning-based human behavior modeling, often at the level of characterizing an entire clinical encounter such as a therapy session, has been shown to be useful across a range of domains in psychological research and practice from relationship and family studies to cancer care. Existing approaches typically first quantify the target behavior construct based on cues in an observation window, such as a fixed number of words, and then aggregate it over all the windows in that session. During this process, a sufficiently long window is employed so that adequate information is gathered to accurately estimate the construct. The link between behavior modeling and the observation length, however, has not been well studied, especially for spoken language. In this paper, we analyze the effect of observation window length on the quality of behavior quantification and present a framework for determining appropriate windows for a wide range of behaviors. Our analysis method employs two levels of evaluations: (a) extrinsic similarity between machine predictions and human expert annotations, and (b) intrinsic consistency between intra-machine and intra-human behavior relations. We apply our analysis on a dataset of real-life married couple interactions that are annotated for a large and diverse set of behavior codes and test the robustness of our findings to different machine learning models. We find that negative constructs such as blame can be accurately identified from short expressions while those pertaining to positive affect such as satisfaction tend to require slightly longer observation windows. Behaviors that describe more complex personality traits such as negotiation and avoidance are found to require very long observations and are difficult to quantify from language alone. Our findings are in agreement with similar work on acoustic cues, thin slices and human emotion perception.
Automatic quantification of human interaction behaviors based on language information has been shown to be effective in psychotherapy research domains such as marital therapy and cancer care. Existing systems typically use a moving-window approach where the target behavior construct is first quantified based on observations inside a window, such as a fixed number of words or turns, and then integrated over all the windows in that interaction. Given a behavior of interest, it is important to employ the appropriate length of observation, since too short a window might not contain sufficient information. Unfortunately, the link between behavior and observation length for lexical cues has not been well studied and it is not clear how these requirements relate to the characteristics of the target behavior construct. Therefore, in this paper, we investigate how the choice of window length affects the efficacy of language-based behavior quantification, by analyzing (a) the similarity between system predictions and human expert assessments for the same behavior construct and (b) the consistency in relations between predictions of related behavior constructs. We apply our analysis to a large and diverse set of behavior codes that are used to annotate real-life interactions and find that behaviors related to negative affect can be quantified from just a few words whereas those related to positive traits and problem solving require much longer observation windows. On the other hand, constructs that describe dysphoric affect do not appear to be quantifiable from language information alone, regardless of how long they are observed. We compare our findings with related work on behavior quantification based on acoustic vocal cues as well as with prior work on thin slices and human personality predictions and find that, in general, they are in agreement.
The ubiquity of mobile devices and wearable sensors offers unprecedented opportunities for continuous collection of multimodal physiological data. Such data enables temporal characterization of an individual's behaviors, which can provide unique insights into her physical and psychological health. Understanding the relation between different behaviors/activities and personality traits such as stress or work performance can help build strategies to improve the work environment. Especially in workplaces like hospitals where many employees are overworked, having such policies improves the quality of patient care by prioritizing mental and physical health of their caregivers. One challenge in analyzing physiological data is extracting the underlying behavioral states from the temporal sensor signals and interpreting them. Here, we use a non-parametric Bayesian approach, to model multivariate sensor data from multiple people and discover dynamic behaviors they share. We apply this method to data collected from sensors worn by a population of workers in a large urban hospital, capturing their physiological signals, such as breathing and heart rate, and activity patterns. We show that the learned states capture behavioral differences within the population that can help cluster participants into meaningful groups and better predict their cognitive and affective states. This method offers a practical way to learn compact behavioral representations from dynamic multivariate sensor signals and provide insights into the data.
Devices capable of detecting and categorizing acoustic scenes have numerous applications such as providing context-aware user experiences. In this paper, we address the task of characterizing acoustic scenes in a workplace setting from audio recordings collected with wearable microphones. The acoustic scenes, tracked with Bluetooth transceivers, vary dynamically with time from the egocentric perspective of a mobile user. Our dataset contains experience sampled long audio recordings collected from clinical providers in a hospital, who wore the audio badges during multiple work shifts. To handle the long egocentric recordings, we propose a Time Delay Neural Network~(TDNN)-based segment-level modeling. The experiments show that TDNN outperforms other models in the acoustic scene classification task. We investigate the effect of primary speaker's speech in determining acoustic scenes from audio badges, and provide a comparison between performance of different models. Moreover, we explore the relationship between the sequence of acoustic scenes experienced by the users and the nature of their jobs, and find that the scene sequence predicted by our model tend to possess similar relationship. The initial promising results reveal numerous research directions for acoustic scene classification via wearable devices as well as egocentric analysis of dynamic acoustic scenes encountered by the users.
Representation learning for speech emotion recognition is challenging due to labeled data sparsity issue and lack of gold standard references. In addition, there is much variability from input speech signals, human subjective perception of the signals and emotion label ambiguity. In this paper, we propose a machine learning framework to obtain speech emotion representations by limiting the effect of speaker variability in the speech signals. Specifically, we propose to disentangle the speaker characteristics from emotion through an adversarial training network in order to better represent emotion. Our method combines the gradient reversal technique with an entropy loss function to remove such speaker information. Our approach is evaluated on both IEMOCAP and CMU-MOSEI datasets. We show that our method improves speech emotion classification and increases generalization to unseen speakers.
Diagnostic procedures for ASD (autism spectrum disorder) involve semi-naturalistic interactions between the child and a clinician. Computational methods to analyze these sessions require an end-to-end speech and language processing pipeline that go from raw audio to clinically-meaningful behavioral features. An important component of this pipeline is the ability to automatically detect who is speaking when i.e., perform child-adult speaker classification. This binary classification task is often confounded due to variability associated with the participants' speech and background conditions. Further, scarcity of training data often restricts direct application of conventional deep learning methods. In this work, we address two major sources of variability - age of the child and data source collection location - using domain adversarial learning which does not require labeled target domain data. We use two methods, generative adversarial training with inverted label loss and gradient reversal layer to learn speaker embeddings invariant to the above sources of variability, and analyze different conditions under which the proposed techniques improve over conventional learning methods. Using a large corpus of ADOS-2 (autism diagnostic observation schedule, 2nd edition) sessions, we demonstrate upto 13.45% and 6.44% relative improvements over conventional learning methods.