arXiv subjects

Wen Ding

Publications and source records attributed to Wen Ding.

12 recordsLinked to original sources

arXiv2025

LESS: Large Language Model Enhanced Semi-Supervised Learning for Speech Foundational Models Using in-the-wild Data

Although state-of-the-art Speech Foundational Models can produce high-quality text pseudo-labels, applying Semi-Supervised Learning (SSL) for in-the-wild real-world data remains challenging due to its richer and more complex acoustics compared to curated datasets. To address the challenges, we introduce LESS (Large Language Model Enhanced Semi-supervised Learning), a versatile framework that uses Large Language Models (LLMs) to correct pseudo-labels generated on in-the-wild data. In the LESS framework, pseudo-labeled text from Automatic Speech Recognition (ASR) or Automatic Speech Translation (AST) of the unsupervised data is refined by an LLM, and further improved by a data filtering strategy. Across Mandarin ASR and Spanish-to-English AST evaluations, LESS delivers consistent gains, with an absolute Word Error Rate reduction of 3.8% on WenetSpeech, and BLEU score increase of 0.8 and 0.7, achieving 34.0 on Callhome and 64.7 on Fisher testsets respectively. These results highlight LESS's effectiveness across diverse languages, tasks, and domains. We have released the recipe as open source to facilitate further research in this area.

cs.CL↗

arXiv2025

CTLformer: A Hybrid Denoising Model Combining Convolutional Layers and Self-Attention for Enhanced CT Image Reconstruction

Low-dose CT (LDCT) images are often accompanied by significant noise, which negatively impacts image quality and subsequent diagnostic accuracy. To address the challenges of multi-scale feature fusion and diverse noise distribution patterns in LDCT denoising, this paper introduces an innovative model, CTLformer, which combines convolutional structures with transformer architecture. Two key innovations are proposed: a multi-scale attention mechanism and a dynamic attention control mechanism. The multi-scale attention mechanism, implemented through the Token2Token mechanism and self-attention interaction modules, effectively captures both fine details and global structures at different scales, enhancing relevant features and suppressing noise. The dynamic attention control mechanism adapts the attention distribution based on the noise characteristics of the input image, focusing on high-noise regions while preserving details in low-noise areas, thereby enhancing robustness and improving denoising performance. Furthermore, CTLformer integrates convolutional layers for efficient feature extraction and uses overlapping inference to mitigate boundary artifacts, further strengthening its denoising capability. Experimental results on the 2016 National Institutes of Health AAPM Mayo Clinic LDCT Challenge dataset demonstrate that CTLformer significantly outperforms existing methods in both denoising performance and model efficiency, greatly improving the quality of LDCT images. The proposed CTLformer not only provides an efficient solution for LDCT denoising but also shows broad potential in medical image analysis, especially for clinical applications dealing with complex noise patterns.

eess.IV↗

arXiv2025

Financial Analysis: Intelligent Financial Data Analysis System Based on LLM-RAG

In the modern financial sector, the exponential growth of data has made efficient and accurate financial data analysis increasingly crucial. Traditional methods, such as statistical analysis and rule-based systems, often struggle to process and derive meaningful insights from complex financial information effectively. These conventional approaches face inherent limitations in handling unstructured data, capturing intricate market patterns, and adapting to rapidly evolving financial contexts, resulting in reduced accuracy and delayed decision-making processes. To address these challenges, this paper presents an intelligent financial data analysis system that integrates Large Language Models (LLMs) with Retrieval-Augmented Generation (RAG) technology. Our system incorporates three key components: a specialized preprocessing module for financial data standardization, an efficient vector-based storage and retrieval system, and a RAG-enhanced query processing module. Using the NASDAQ financial fundamentals dataset from 2010 to 2023, we conducted comprehensive experiments to evaluate system performance. Results demonstrate significant improvements across multiple metrics: the fully optimized configuration (gpt-3.5-turbo-1106+RAG) achieved 78.6% accuracy and 89.2% recall, surpassing the baseline model by 23 percentage points in accuracy while reducing response time by 34.8%. The system also showed enhanced efficiency in handling complex financial queries, though with a moderate increase in memory utilization. Our findings validate the effectiveness of integrating RAG technology with LLMs for financial analysis tasks and provide valuable insights for future developments in intelligent financial data processing systems.

q-fin.ST↗

arXiv2024

NTC-KWS: Noise-aware CTC for Robust Keyword Spotting

In recent years, there has been a growing interest in designing small-footprint yet effective Connectionist Temporal Classification based keyword spotting (CTC-KWS) systems. They are typically deployed on low-resource computing platforms, where limitations on model size and computational capacity create bottlenecks under complicated acoustic scenarios. Such constraints often result in overfitting and confusion between keywords and background noise, leading to high false alarms. To address these issues, we propose a noise-aware CTC-based KWS (NTC-KWS) framework designed to enhance model robustness in noisy environments, particularly under extremely low signal-to-noise ratios. Our approach introduces two additional noise-modeling wildcard arcs into the training and decoding processes based on weighted finite state transducer (WFST) graphs: self-loop arcs to address noise insertion errors and bypass arcs to handle masking and interference caused by excessive noise. Experiments on clean and noisy Hey Snips show that NTC-KWS outperforms state-of-the-art (SOTA) end-to-end systems and CTC-KWS baselines across various acoustic conditions, with particularly strong performance in low SNR scenarios.

eess.AS↗

arXiv2024

Semi-supervised Learning for Code-Switching ASR with Large Language Model Filter

Code-switching (CS) phenomenon occurs when words or phrases from different languages are alternated in a single sentence. Due to data scarcity, building an effective CS Automatic Speech Recognition (ASR) system remains challenging. In this paper, we propose to enhance CS-ASR systems by utilizing rich unsupervised monolingual speech data within a semi-supervised learning framework, particularly when access to CS data is limited. To achieve this, we establish a general paradigm for applying noisy student training (NST) to the CS-ASR task. Specifically, we introduce the LLM-Filter, which leverages well-designed prompt templates to activate the correction capability of large language models (LLMs) for monolingual data selection and pseudo-labels refinement during NST. Our experiments on the supervised ASRU-CS and unsupervised AISHELL-2 and LibriSpeech datasets show that our method not only achieves significant improvements over supervised and semi-supervised learning baselines for the CS task, but also attains better performance compared with the fully-supervised oracle upper-bound on the CS English part. Additionally, we further investigate the influence of accent on AESRC dataset and demonstrate that our method can get achieve additional benefits when the monolingual data contains relevant linguistic characteristic.

eess.AS↗

arXiv2024

Romanization Encoding For Multilingual ASR

We introduce romanization encoding for script-heavy languages to optimize multilingual and code-switching Automatic Speech Recognition (ASR) systems. By adopting romanization encoding alongside a balanced concatenated tokenizer within a FastConformer-RNNT framework equipped with a Roman2Char module, we significantly reduce vocabulary and output dimensions, enabling larger training batches and reduced memory consumption. Our method decouples acoustic modeling and language modeling, enhancing the flexibility and adaptability of the system. In our study, applying this method to Mandarin-English ASR resulted in a remarkable 63.51% vocabulary reduction and notable performance gains of 13.72% and 15.03% on SEAME code-switching benchmarks. Ablation studies on Mandarin-Korean and Mandarin-Japanese highlight our method's strong capability to address the complexities of other script-heavy languages, paving the way for more versatile and effective multilingual ASR systems.

cs.CL↗

arXiv2024

Gecko: Versatile Text Embeddings Distilled from Large Language Models

We present Gecko, a compact and versatile text embedding model. Gecko achieves strong retrieval performance by leveraging a key idea: distilling knowledge from large language models (LLMs) into a retriever. Our two-step distillation process begins with generating diverse, synthetic paired data using an LLM. Next, we further refine the data quality by retrieving a set of candidate passages for each query, and relabeling the positive and hard negative passages using the same LLM. The effectiveness of our approach is demonstrated by the compactness of the Gecko. On the Massive Text Embedding Benchmark (MTEB), Gecko with 256 embedding dimensions outperforms all existing entries with 768 embedding size. Gecko with 768 embedding dimensions achieves an average score of 66.31, competing with 7x larger models and 5x higher dimensional embeddings.

cs.CL↗

arXiv2023

Wespeaker baselines for VoxSRC2023

This report showcases the results achieved using the wespeaker toolkit for the VoxSRC2023 Challenge. Our aim is to provide participants, especially those with limited experience, with clear and straightforward guidelines to develop their initial systems. Via well-structured recipes and strong results, we hope to offer an accessible and good enough start point for all interested individuals. In this report, we describe the results achieved on the VoxSRC2023 dev set using the pretrained models, you can check the CodaLab evaluation server for the results on the evaluation set.

eess.AS↗

arXiv2022

Improving Noisy Student Training on Non-target Domain Data for Automatic Speech Recognition

Noisy Student Training (NST) has recently demonstrated extremely strong performance in Automatic Speech Recognition(ASR). In this paper, we propose a data selection strategy named LM Filter to improve the performance of NST on non-target domain data in ASR tasks. Hypotheses with and without a Language Model are generated and the CER differences between them are utilized as a filter threshold. Results reveal that significant improvements of 10.4% compared with no data filtering baselines. We can achieve 3.31% CER in AISHELL-1 test set, which is best result from our knowledge without any other supervised data. We also perform evaluations on the supervised 1000 hour AISHELL-2 dataset and competitive results of 4.73% CER can be achieved.

cs.SD↗

arXiv2022

Coupling Deep Learning with Full Waveform Inversion

Full waveform inversion (FWI) aims to reconstruct unknown physical coefficients in wave equations using the wavefield data generated from multiple incoming sources. In this work, we propose an offline-online computational strategy for coupling classical least-squares-based computational inversion with modern learning-based approaches for FWI to achieve advantages that can not be achieved with only one of the components. \RED{In brief, we develop an offline learning strategy to construct a robust approximation of the inverse operator through weighted optimization and utilize it to design a new objective function for approximate online inversion with new datasets. The approximate online inversion then serves as a warm start for the true online inversion.} We demonstrate through numerical simulations that our coupling strategy improves the computational efficiency of FWI with reliable offline training on moderate computational resources (in terms of both the size of the training dataset and the computational cost needed).

math.NA↗

arXiv2021

Deep Neural Networks for Creating Reliable PmP Database with a Case Study in Southern California

Recent progresses in artificial intelligence and machine learning make it possible to automatically identify seismic phases from exponentially growing seismic data. Despite some exciting successes in automatic picking of the first P- and S-wave arrivals, auto-identification of later seismic phases such as the Moho-reflected PmP waves remains a significant challenge in matching the performance of experienced analysts. The main difficulty of machine-identifying PmP waves is that the identifiable PmP waves are rare, making the problem of identifying the PmP waves from a massive seismic database inherently unbalanced. In this work, by utilizing a high-quality PmP dataset (10,192 manual picks) in southern California, we develop PmPNet, a deep-neural-network-based algorithm to automatically identify PmP waves efficiently; by doing so, we accelerate the process of identifying the PmP waves. PmPNet applies similar techniques in the machine learning community to address the unbalancement of PmP datasets. The architecture of PmPNet is a residual neural network (ResNet)-autoencoder with additional predictor block, where encoder, decoder, and predictor are equipped with ResNet connection. We conduct systematic research with field data, concluding that PmPNet can efficiently achieve high precision and high recall simultaneously to automatically identify PmP waves from a massive seismic database. Applying the pre-trained PmPNet to the seismic database from January 1990 to December 1999 in southern California, we obtain nearly twice more PmP picks than the original PmP dataset, providing valuable data for other studies such as mapping the topography of the Moho discontinuity and imaging the lower crust structures of southern California.

physics.geo-ph↗

arXiv2005

A Game Theoretic Economics Framework to understanding Information Security Oursourcing Market

On information security outsourcing market, an important reason that firms do not want to let outside firms(usually called MSSPs-Managed Security Service Providers) to take care of their security need is that they worry about service quality MSSPs provide because they cannot monitor effort of the MSSPs. Since MSSPs action is unobservable to buyers, MSSPs can lower cost by working less hard than required in the contract and get higher profit. In the asymmetric information literature, this possible secret shirking behavior is termed as moral hazard problem. This paper considers a game theoretic economic framework to show that under information asymmetry, an optimal contract can be designed so that MSSPs will stick to their promised effort level. We also show that the optimal contract should be performance-based, i.e., payment to MSSP should base on performance of MSSP's security service period by period. For comparison, we also showed that if the moral hazard problem does not exist, the optimal contract does not depend on MSSP's performance. A contract that specifies constant payment to MSSP will be optimal. Besides these, we show that for no matter under perfect information scenario or imperfect information scenario, the higher the transaction cost is, the lower payment to MSSPs will be.

cs.GT↗