An official website of the United States government
Abstract In this article, we present a live speech-driven, avatar-mediated, three-party telepresence system, through which three distant users, embodied as avatars in a shared 3D virtual world, can perform natural three-party telepresence that does not require tracking devices. Based on live speech input from three users, this system can real-time generate the corresponding conversational motions of all the avatars, including head motion, eye motion, lip movement, torso motion, and hand gesture. All motions are generated automatically at each user side based on live speech input, and a cloud server is utilized to transmit and synchronize motion and speech among different users. We conduct a formal user study to evaluate the usability and effectiveness of the system by comparing it with a well-known online virtual world, Second Life, and a widely-used online teleconferencing system, Skype. The user study results indicate our system can provide a measurably better telepresence user experience than the two widely-used methods.
more »
« less
S2M-Net: Speech Driven Three-party Conversational Motion Synthesis Networks
In this paper we propose a novel conditional generative adversarial network (cGAN) architecture, called S2M-Net, to holistically synthesize realistic three-party conversational animations based on acoustic speech input together with speaker marking (i.e., the speak- ing time of each interlocutor). Specifically, based on a pre-collected three-party conversational motion dataset, we design and train the S2M-Net for three-party conversational animation synthesis. In the architecture, a generator contains a LSTM encoder to encode a sequence of acoustic speech features to a latent vector that is further fed into a transform unit to transform the latent vector into a gesture kinematics space. Then, the output of this transform unit is fed into a LSTM decoder to generate corresponding three-party conversational gesture kinematics. Meanwhile, a discriminator is implemented to check whether an input sequence of three-party conversational gesture kinematics is real or fake. To evaluate our method, besides quantitative and qualitative evaluations, we also conducted paired comparison user studies to compare it with the state of the art.
more »
« less
- Award ID(s):
- 2005430
- PAR ID:
- 10463603
- Date Published:
- Journal Name:
- Proceeding of ACM SIGGRAPH Conference on Motion, Interaction, and Games 2022
- Page Range / eLocation ID:
- 2:1 to 2:10
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Theunissen, Frédéric E (Ed.)Human speech recognition transforms a continuous acoustic signal into categorical linguistic units, by aggregating information that is distributed in time. It has been suggested that this kind of information processing may be understood through the computations of a Recurrent Neural Network (RNN) that receives input frame by frame, linearly in time, but builds an incremental representation of this input through a continually evolving internal state. While RNNs can simulate several keybehavioralobservations about human speech and language processing, it is unknown whether RNNs also develop computational dynamics that resemble humanneural speech processing. Here we show that the internal dynamics of long short-term memory (LSTM) RNNs, trained to recognize speech from auditory spectrograms, predict human neural population responses to the same stimuli, beyond predictions from auditory features. Variations in the RNN architecture motivated by cognitive principles further improved this predictive power. Specifically, modifications that allow more human-like phonetic competition also led to more human-like temporal dynamics. Overall, our results suggest that RNNs provide plausible computational models of the cortical processes supporting human speech recognition.more » « less
-
Speaker Verification (SV) systems trained on adults speech often underperform on children’s SV due to the acoustic mismatch, and limited children speech data makes fine-tuning not very effective. In this paper, we propose an innovative framework, a Gated Linear Unit adapter with Iterative Fine-Tuning (G-IFT), to enhance knowledge transfer efficiency between the high-resource adults speech domain and the low-resource chil- dren’s speech domain. In this framework, a Gated Linear Unit adapter is first inserted between the pre-trained speaker embedding model and the classifier. Then the classifier, adapter, and pre-trained speaker embedding model are optimized sequentially in an iterative way. This framework is agnostic to the type of the underlying architecture of the SV system. Our experiments on ECAPA-TDNN, ResNet, and X-vector architectures using the OGI and MyST datasets demonstrate that the G-IFT framework yields consistent reductions in Equal Error Rates compared to baseline methods.more » « less
-
ISCA (Ed.)In this paper, we present MixRep, a simple and effective data augmentation strategy based on mixup for low-resource ASR. MixRep interpolates the feature dimensions of hidden representations in the neural network that can be applied to both the acoustic feature input and the output of each layer, which generalizes the previous MixSpeech method. Further, we propose to combine the mixup with a regularization along the time axis of the input, which is shown as complementary. We apply MixRep to a Conformer encoder of an E2E LAS architecture trained with a joint CTC loss. We experiment on the WSJ dataset and subsets of the SWB dataset, covering reading and telephony conversational speech. Experimental results show that MixRep consistently outperforms other regularization methods for low-resource ASR. Compared to a strong SpecAugment baseline, MixRep achieves a +6.5% and a +6.7% relative WER reduction on the eval92 set and the Callhome part of the eval'2000 set.more » « less
-
Abstract Photospheric magnetic field parameters are frequently used to analyze and predict solar events. Observation of these parameters over time, i.e., representing solar events by multivariate time-series (MVTS) data, can determine relationships between magnetic field states in active regions and extreme solar events, e.g., solar flares. We can improve our understanding of these events by selecting the most relevant parameters that give the highest predictive performance. In this study, we propose a two-step incremental feature selection method for MVTS data using a deep-learning model based on long short-term memory (LSTM) networks. First, each MVTS feature (magnetic field parameter) is evaluated individually by a univariate sequence classifier utilizing an LSTM network. Then, the top performing features are combined to produce input for an LSTM-based multivariate sequence classifier. Finally, we tested the discrimination ability of the selected features by training downstream classifiers, e.g., Minimally Random Convolutional Kernel Transform and support vector machine. We performed our experiments using a benchmark data set for flare prediction known as Space Weather Analytics for Solar Flares. We compared our proposed method with three other baseline feature selection methods and demonstrated that our method selects more discriminatory features compared to other methods. Due to the imbalanced nature of the data, primarily caused by the rarity of minority flare classes (e.g., the X and M classes), we used the true skill statistic as the evaluation metric. Finally, we reported the set of photospheric magnetic field parameters that give the highest discrimination performance in predicting flare classes.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript
- Conference Paper:
- https://doi.org/10.1145/3561975.3562954
