Selected work

Representative Papers

ADI-20: Arabic Dialect Identification dataset and models

Aug 17, 2025Interspeech

This work addresses Arabic Dialect Identification (ADI), a challenging multilingual speech classification task. We introduce ADI-20, the first large-scale, publicly available dataset covering all 22 Arab League countries’ dialects plus Modern Standard Arabic (MSA), comprising 19 dialects and 3,556 hours of speech. We propose an end-to-end ADI model built upon the ECAPA-TDNN backbone, enhanced with Whisper encoder blocks, attention-based pooling, and a dialect-specific classification head. To our knowledge, this is the first open, reproducible framework for pan-Arabic dialect modeling—releasing data, models, and training code. Experiments demonstrate state-of-the-art performance, exceptional data efficiency (F1 drops by <1.5% when trained on only 30% of the data), and systematic empirical analysis of the scaling relationships between dataset size, model parameters, and ADI accuracy.

2 citationsRead paper

ELYADATA & LIA at NADI 2025: ASR and ADI Subtasks

Nov 13, 2025Proceedings of The Third Arabic Natural Language Processing Conference: Shared Tasks

This work addresses the NADI multilingual Arabic dialect speech processing task, tackling two core challenges: Arabic Dialect Identification (ADI) and multilingual Automatic Speech Recognition (ASR). We propose a joint optimization framework based on large-model fine-tuning and dialect-specific data augmentation. For ADI, we employ the Whisper-large-v3 encoder with dialect-aware data augmentation to achieve end-to-end dialect classification. For ASR, we fine-tune the SeamlessM4T-v2 Large model separately on each of eight Arabic dialects to enhance cross-dialect robustness. Our approach significantly outperforms baselines: achieving 79.83% accuracy on ADI (ranked first), and average WER/CER of 38.54%/14.53% on ASR (ranked second). The key contribution lies in empirically validating that dialect-specific fine-tuning combined with domain-adaptive data augmentation substantially improves low-resource multilingual speech modeling performance.

1 citations1 influentialRead paper

FedSV: Byzantine-Robust Federated Learning via Shapley Value

Jun 09, 2024ICC 2024 - IEEE International Conference on Communications

To address the degradation of global model performance in federated learning caused by malicious client poisoning attacks, this paper proposes FedSV, a robust aggregation mechanism based on Shapley values. FedSV’s core innovation lies in the first extension of Shapley value computation to estimate *multi-group collaborative contributions*, dynamically quantifying each client’s marginal contribution across diverse client subsets—enabling fine-grained, context-aware detection of anomalous behavior. By jointly modeling accuracy-based marginal gains and optimizing federated aggregation, FedSV is specifically designed for cross-silo settings. Extensive experiments on MNIST under various Byzantine attack scenarios demonstrate that FedSV improves global model accuracy by up to 12.7% and achieves an F1-score exceeding 0.93 for malicious client identification—substantially enhancing system robustness against adversarial clients.

1 citationsRead paper

Why Do You Say It Like That? A Phoneme-Level Framework for Explainable Speech Deepfake Detection

Jul 09, 2026

This work addresses the limited interpretability of current deepfake speech detection models, which often fail to reveal the basis for their decisions. To enhance transparency, the authors propose a phoneme-level interpretable analysis framework that integrates Grad-CAM with automatic speech recognition, leveraging self-supervised speech representations from models such as wav2vec 2.0 or HuBERT to generate phoneme-aligned saliency maps. These maps effectively visualize acoustic anomalies in spoofed speech. Evaluated on the ASVspoof 5 dataset, the method achieves detection performance comparable to state-of-the-art models while uncovering statistically significant, attack-type- and speaker-dependent phoneme-level artifacts. This advancement substantially improves the interpretability and trustworthiness of deepfake detection systems by providing fine-grained, linguistically grounded explanations for model predictions.

0 citationsRead paper

Streaming Neural Speech Codecs through Time-Invariant Representations

Jul 06, 2026

This work addresses the challenges faced by conventional neural speech codecs in low-latency scenarios, where heavy frame-level modeling burdens hinder the simultaneous optimization of reconstruction quality and computational efficiency. To overcome this, the authors propose TiCodec, a novel framework featuring a Time-Invariant Representation Extraction (TIRE) module that disentangles speech into time-varying and time-invariant components, substantially reducing frame-level modeling complexity while enabling streaming processing. By incorporating a Dual-TIRE multi-layer architecture that fuses complementary information from different encoder depths, along with factorized representation learning and a 660-ms chunk-based streaming inference strategy, the method significantly enhances both reconstruction fidelity and speaker similarity. Experimental results demonstrate that TiCodec achieves near non-streaming performance under streaming conditions, making it well-suited for low-latency speech generation systems.

0 citationsRead paper

Recent publications

Latest Papers

Why Do You Say It Like That? A Phoneme-Level Framework for Explainable Speech Deepfake Detection

Jul 09, 2026

This work addresses the limited interpretability of current deepfake speech detection models, which often fail to reveal the basis for their decisions. To enhance transparency, the authors propose a phoneme-level interpretable analysis framework that integrates Grad-CAM with automatic speech recognition, leveraging self-supervised speech representations from models such as wav2vec 2.0 or HuBERT to generate phoneme-aligned saliency maps. These maps effectively visualize acoustic anomalies in spoofed speech. Evaluated on the ASVspoof 5 dataset, the method achieves detection performance comparable to state-of-the-art models while uncovering statistically significant, attack-type- and speaker-dependent phoneme-level artifacts. This advancement substantially improves the interpretability and trustworthiness of deepfake detection systems by providing fine-grained, linguistically grounded explanations for model predictions.

0 citationsRead paper

Streaming Neural Speech Codecs through Time-Invariant Representations

Jul 06, 2026

This work addresses the challenges faced by conventional neural speech codecs in low-latency scenarios, where heavy frame-level modeling burdens hinder the simultaneous optimization of reconstruction quality and computational efficiency. To overcome this, the authors propose TiCodec, a novel framework featuring a Time-Invariant Representation Extraction (TIRE) module that disentangles speech into time-varying and time-invariant components, substantially reducing frame-level modeling complexity while enabling streaming processing. By incorporating a Dual-TIRE multi-layer architecture that fuses complementary information from different encoder depths, along with factorized representation learning and a 660-ms chunk-based streaming inference strategy, the method significantly enhances both reconstruction fidelity and speaker similarity. Experimental results demonstrate that TiCodec achieves near non-streaming performance under streaming conditions, making it well-suited for low-latency speech generation systems.

0 citationsRead paper

The Capacitated $p$-Location Problem with Territorial Coverage Constraint: Efficient Solution and Case Study

Jul 02, 2026

This study addresses the capacitated p-location problem by introducing, for the first time, multi-scale area coverage constraints to ensure spatial fairness. To this end, the authors propose two models—C$p$LP-TC and its multi-scale extension C$p$LP-MTC—along with a strengthened integer linear programming (ILP) formulation, valid inequalities, and a parameter-free randomized sampling spatial voting (RSSV) heuristic, integrated with problem-size reduction techniques for efficient solution. Evaluation on an open-source benchmark dataset constructed from French administrative regions demonstrates that the approach effectively quantifies the trade-off between service efficiency and regional equity, offering scalable and reproducible decision support for fair facility location in real-world applications.

0 citationsRead paper

Linguistic Bias Mitigation for Spoofing Detection via Gradient Reversal and A Variational Information Bottleneck

Jun 30, 2026

This study addresses the poor generalization of existing voice anti-spoofing models in cross-domain scenarios, which primarily stems from their reliance on linguistic content—referred to as language bias. The work identifies language bias for the first time as a key factor degrading cross-domain performance and proposes a language-invariant detection framework. This framework leverages a pretrained language-aware teacher model to guide a student model, integrating a gradient reversal layer with a variational information bottleneck to suppress language-related cues while preserving discriminative acoustic features. Evaluated on nine DF-Audio datasets, the proposed method significantly outperforms baseline approaches, achieving a relative reduction in equal error rate (EER) of up to 36.2%.

0 citationsRead paper

An LLM-Based Framework for Intent-Driven Network Topology Design

Jun 30, 2026

This work addresses the core challenge in network automation: automatically generating deployable network topologies from natural language requirements while satisfying structural and resilience constraints. We propose a large language model (LLM)-based, constraint-driven framework that translates natural language into compliant topologies through hierarchical intent parsing and systematic validation. To facilitate evaluation, we introduce the first benchmark for this task, releasing a public dataset encompassing four real-world scenarios and characterizing common generation error patterns. Extensive experiments across multiple proprietary and open-source LLMs demonstrate the framework’s effectiveness, with performance quantified using metrics including topological correctness, node/edge F1 scores, and server-content connectivity. Our results provide actionable guidance for model selection in AI-driven network design.

0 citationsRead paper