We introduce a new cross-modal fusion technique designed for generative error correction in automatic speech recognition (ASR). Our methodology leverages both acoustic information and external linguistic representations to generate accurate speech transcription contexts. This marks a step towards a fresh paradigm in generative error correction within the realm of n-best hypotheses. Unlike the existing ranking-based rescoring methods, our approach adeptly uses distinct initialization techniques and parameter-efficient algorithms to boost ASR performance derived from pre-trained speech and text models. Through evaluation across diverse ASR datasets, we evaluate the stability and reproducibility of our fusion technique, demonstrating its improved word error rate relative (WERR) performance in comparison to n-best hypotheses by relatively 37.66%. To encourage future research, we have made our code and pre-trained models open source at https://github.com/Srijith-rkr/Whispering-LLaMA.
Large language models offer new ways of empowering people to program robot applications-namely, code generation via prompting. However, the code generated by LLMs is susceptible to errors. This work reports a preliminary exploration that empirically characterizes common errors produced by LLMs in robot programming. We categorize these errors into two phases: interpretation and execution. In this work, we focus on errors in execution and observe that they are caused by LLMs being "forgetful" of key information provided in user prompts. Based on this observation, we propose prompt engineering tactics designed to reduce errors in execution. We then demonstrate the effectiveness of these tactics with three language models: ChatGPT, Bard, and LLaMA-2. Finally, we discuss lessons learned from using LLMs in robot programming and call for the benchmarking of LLM-powered end-user development of robot applications.
This paper proposes a simple and robust zero-shot voice conversion system with a cycle structure and mel-spectrogram pre-processing. Previous works suffer from information loss and poor synthesis quality due to their reliance on a carefully designed bottleneck structure. Moreover, models relying solely on self-reconstruction loss struggled with reproducing different speakers' voices. To address these issues, we suggested a cycle-consistency loss that considers conversion back and forth between target and source speakers. Additionally, stacked random-shuffled mel-spectrograms and a label smoothing method are utilized during speaker encoder training to extract a time-independent global speaker representation from speech, which is the key to a zero-shot conversion. Our model outperforms existing state-of-the-art results in both subjective and objective evaluations. Furthermore, it facilitates cross-lingual voice conversions and enhances the quality of synthesized speech.
Annotators' sociodemographic backgrounds (i.e., the individual compositions of their gender, age, educational background, etc.) have a strong impact on their decisions when working on subjective NLP tasks, such as hate speech detection. Often, heterogeneous backgrounds result in high disagreements. To model this variation, recent work has explored sociodemographic prompting, a technique, which steers the output of prompt-based models towards answers that humans with specific sociodemographic profiles would give. However, the available NLP literature disagrees on the efficacy of this technique -- it remains unclear, for which tasks and scenarios it can help and evaluations are limited to specific tasks only. We address this research gap by presenting the largest and most comprehensive study of sociodemographic prompting today. Concretely, we evaluate several prompt formulations across seven datasets and six instruction-tuned model families. We find that (1) while sociodemographic prompting can be beneficial for improving zero-shot learning in subjective NLP tasks, (2) its outcomes largely vary for different model types, sizes, and datasets, (3) are subject to large variance with regards to prompt formulations. Thus, sociodemographic prompting is not a reliable proxy for traditional data annotation with a sociodemographically heterogeneous group of annotators. Instead, we propose (4) to use it for identifying ambiguous instances resulting in more informed annotation efforts.
We introduced Temporally Incremental Disparity Estimation Network (TIDE-Net), a learning-based technique for disparity computation in mono-camera structured light systems. In our hardware setting, a static pattern is projected onto a dynamic scene and captured by a monocular camera. Different from most former disparity estimation methods that operate in a frame-wise manner, our network acquires disparity maps in a temporally incremental way. Specifically, We exploit the deformation of projected patterns (named pattern flow ) on captured image sequences, to model the temporal information. Notably, this newly proposed pattern flow formulation reflects the disparity changes along the epipolar line, which is a special form of optical flow. Tailored for pattern flow, the TIDE-Net, a recurrent architecture, is proposed and implemented. For each incoming frame, our model fuses correlation volumes (from current frame) and disparity (from former frame) warped by pattern flow. From fused features, the final stage of TIDE-Net estimates the residual disparity rather than the full disparity, as conducted by many previous methods. Interestingly, this design brings clear empirical advantages in terms of efficiency and generalization ability. Using only synthetic data for training, our extensitve evaluation results (w.r.t. both accuracy and efficienty metrics) show superior performance than several SOTA models on unseen real data. The code is available on https://github.com/CodePointer/TIDENet.
Salient object detection (SOD) in panoramic video is still in the initial exploration stage. The indirect application of 2D video SOD method to the detection of salient objects in panoramic video has many unmet challenges, such as low detection accuracy, high model complexity, and poor generalization performance. To overcome these hurdles, we design an Inter-Layer Attention (ILA) module, an Inter-Layer weight (ILW) module, and a Bi-Modal Attention (BMA) module. Based on these modules, we propose a Spatial-Temporal Dual-Mode Mixed Flow Network (STDMMF-Net) that exploits the spatial flow of panoramic video and the corresponding optical flow for SOD. First, the ILA module calculates the attention between adjacent level features of consecutive frames of panoramic video to improve the accuracy of extracting salient object features from the spatial flow. Then, the ILW module quantifies the salient object information contained in the features of each level to improve the fusion efficiency of the features of each level in the mixed flow. Finally, the BMA module improves the detection accuracy of STDMMF-Net. A large number of subjective and objective experimental results testify that the proposed method demonstrates better detection accuracy than the state-of-the-art (SOTA) methods. Moreover, the comprehensive performance of the proposed method is better in terms of memory required for model inference, testing time, complexity, and generalization performance.
Recently, channel extrapolation has been widely investigated in frequency division duplex (FDD) massive MIMO systems. However, in time division duplex (TDD) fifth generation (5G) new radio (NR) systems, the channel extrapolation problem also arises due to the hopping uplink pilot pattern, which has not been fully researched yet. This paper addresses this gap by formulating a channel extrapolation problem in TDD massive MIMO-OFDM systems for 5G NR, incorporating imperfection factors. A novel two-stage two-dimensional (2D) channel extrapolation scheme in both frequency and time domain is proposed, designed to mitigate the negative effects of imperfection factors and ensure high-accuracy channel estimation. Specifically, in the channel estimation stage, we propose a novel multi-band and multi-timeslot based high-resolution parameter estimation algorithm to achieve 2D channel extrapolation in the presence of imperfection factors. Then, to avoid repeated multi-timeslot based channel estimation, a channel tracking stage is designed during the subsequent time instants, in which a sparse Markov channel model is formulated to capture the dynamic sparsity of massive MIMO-OFDM channels under the influence of imperfection factors. Next, an expectation-maximization (EM) based compressive channel tracking algorithm is designed to jointly estimate unknown imperfection and channel parameters by exploiting the high-resolution prior information of the delay/angle parameters from the previous timeslots. Simulation results underscore the superior performance of our proposed channel extrapolation scheme over baselines.
Human activity recognition (HAR) with wearables is one of the serviceable technologies in ubiquitous and mobile computing applications. The sliding-window scheme is widely adopted while suffering from the multi-class windows problem. As a result, there is a growing focus on joint segmentation and recognition with deep-learning methods, aiming at simultaneously dealing with HAR and time-series segmentation issues. However, obtaining the full activity annotations of wearable data sequences is resource-intensive or time-consuming, while unsupervised methods yield poor performance. To address these challenges, we propose a novel method for joint activity segmentation and recognition with timestamp supervision, in which only a single annotated sample is needed in each activity segment. However, the limited information of sparse annotations exacerbates the gap between recognition and segmentation tasks, leading to sub-optimal model performance. Therefore, the prototypes are estimated by class-activation maps to form a sample-to-prototype contrast module for well-structured embeddings. Moreover, with the optimal transport theory, our approach generates the sample-level pseudo-labels that take advantage of unlabeled data between timestamp annotations for further performance improvement. Comprehensive experiments on four public HAR datasets demonstrate that our model trained with timestamp supervision is superior to the state-of-the-art weakly-supervised methods and achieves comparable performance to the fully-supervised approaches.
Existing deep learning-based depth completion methods generally employ massive stacked layers to predict the dense depth map from sparse input data. Although such approaches greatly advance this task, their accompanied huge computational complexity hinders their practical applications. To accomplish depth completion more efficiently, we propose a novel lightweight deep network framework, the Long-short Range Recurrent Updating (LRRU) network. Without learning complex feature representations, LRRU first roughly fills the sparse input to obtain an initial dense depth map, and then iteratively updates it through learned spatially-variant kernels. Our iterative update process is content-adaptive and highly flexible, where the kernel weights are learned by jointly considering the guidance RGB images and the depth map to be updated, and large-to-small kernel scopes are dynamically adjusted to capture long-to-short range dependencies. Our initial depth map has coarse but complete scene depth information, which helps relieve the burden of directly regressing the dense depth from sparse ones, while our proposed method can effectively refine it to an accurate depth map with less learnable parameters and inference time. Experimental results demonstrate that our proposed LRRU variants achieve state-of-the-art performance across different parameter regimes. In particular, the LRRU-Base model outperforms competing approaches on the NYUv2 dataset, and ranks 1st on the KITTI depth completion benchmark at the time of submission. Project page: https://npucvr.github.io/LRRU/.
Deep learning-based fault diagnosis (FD) approaches require a large amount of training data, which are difficult to obtain since they are located across different entities. Federated learning (FL) enables multiple clients to collaboratively train a shared model with data privacy guaranteed. However, the domain discrepancy and data scarcity problems among clients deteriorate the performance of the global FL model. To tackle these issues, we propose a novel framework called representation encoding-based federated meta-learning (REFML) for few-shot FD. First, a novel training strategy based on representation encoding and meta-learning is developed. It harnesses the inherent heterogeneity among training clients, effectively transforming it into an advantage for out-of-distribution generalization on unseen working conditions or equipment types. Additionally, an adaptive interpolation method that calculates the optimal combination of local and global models as the initialization of local training is proposed. This helps to further utilize local information to mitigate the negative effects of domain discrepancy. As a result, high diagnostic accuracy can be achieved on unseen working conditions or equipment types with limited training data. Compared with the state-of-the-art methods, such as FedProx, the proposed REFML framework achieves an increase in accuracy by 2.17%-6.50% when tested on unseen working conditions of the same equipment type and 13.44%-18.33% when tested on totally unseen equipment types, respectively.