Selected work

Representative Papers

FLEURS-Kobani: Extending the FLEURS Dataset for Northern Kurdish

Mar 31, 2026

This study addresses the scarcity of publicly available speech benchmarks for Northern Kurdish (Kurmanji, KMR), which has hindered progress in automatic speech recognition (ASR) and speech translation. To bridge this gap, the authors introduce FLEURS-Kobani, the first KMR-focused dataset derived from the FLEURS benchmark, comprising 5,162 validated utterances (18 hours and 24 minutes) from 31 native speakers, enabling research in ASR, end-to-end speech-to-text translation (S2TT), and speech-to-speech translation (S2ST). Leveraging the Whisper v3-large architecture with Common Voice pretraining and a two-stage fine-tuning strategy, the proposed system achieves a word error rate (WER) of 28.11% and character error rate (CER) of 9.84% on ASR, and a BLEU score of 8.68 for end-to-end KMR→EN S2TT, establishing a foundational benchmark for low-resource Kurdish speech processing.

0 citationsRead paper

SENS-ASR: Semantic Embedding injection in Neural-transducer for Streaming Automatic Speech Recognition

Feb 17, 2026

This work proposes a semantics-enhanced streaming automatic speech recognition (ASR) approach to address performance degradation caused by the absence of future contextual information in low-latency scenarios. By integrating a knowledge distillation–based sentence-embedding language model with a context-aware semantic module, the method extracts high-level semantic cues from historical frames and injects them into a neural Transformer architecture. This integration effectively mitigates the context deficiency inherent in chunk-based streaming processing. Experimental results on standard benchmarks demonstrate a significant reduction in word error rate, validating the efficacy and novelty of leveraging semantic guidance to enhance streaming transcription quality.

0 citationsRead paper

Toward More Reliable Artificial Intelligence: Reducing Hallucinations in Vision-Language Models

Dec 08, 2025

Visual language models (VLMs) suffer from hallucination—generating factually inconsistent or image-ungrounded statements. To address this, we propose a training-free, parameter-free self-correction framework that iteratively refines model outputs via uncertainty-guided visual re-attention. Specifically, it quantifies token-level uncertainty across four dimensions—token entropy, attention dispersion, semantic consistency, and claim confidence—to dynamically identify and crop unreliable image regions, followed by response revision. The method requires no fine-tuning, external data, or architectural modifications. Evaluated on Qwen2.5-VL-7B, it reduces hallucination rate by 9.8 percentage points and improves object existence accuracy by 4.7 percentage points, outperforming existing training-free baselines. Our core contribution lies in the tight coupling of multi-dimensional uncertainty modeling with dynamic visual attention, enabling interpretable, lightweight, and reliable VLM inference without parameter updates.

0 citationsRead paper

Sparse deepfake detection promotes better disentanglement

Oct 07, 2025

To address the trade-off between model interpretability and detection performance in deepfake speech detection, this paper introduces a Top-K sparse activation mechanism into the embedding layer of the AASIST architecture, proposing a latent-space sparse disentangled representation method. Inspired by sparse autoencoders, the approach enforces controllable sparsity to compel the model to focus exclusively on discriminative spoofing features, thereby achieving disentangled encoding of attack patterns in the latent space. To quantitatively evaluate disentanglement quality, we design integrity and modularity metrics based on mutual information. Experimental results on the ASVSpoof2021 LA and DF evaluation sets demonstrate an EER of 23.36% with 95% sparsity—significantly improving both detection accuracy and interpretability while preserving model lightweightness.

0 citationsRead paper

A dependent and censored first hitting-time model with compound Poisson processes

Apr 11, 2025

This paper addresses the challenge of modeling bivariate first-hitting times under dependence and right-censoring—common in clinical dual-endpoint settings such as concurrent liver and kidney injury. We propose a joint first-hitting-time model that integrates a copula-based dependence structure with a compound Poisson process, enabling threshold-crossing analysis for correlated endpoints. Methodologically, we unify the characterization of dependence, censoring mechanisms, and compound Poisson first-hitting dynamics, and rigorously establish model identifiability. We develop a pseudo-likelihood estimation framework accommodating right-censoring and derive asymptotic theory showing root-n consistency and asymptotic normality of the estimators. Monte Carlo simulations confirm robust finite-sample performance. Applied to real-world mushroom poisoning data, our model successfully quantifies temporal dependence between hepatic and renal injury onset and significantly improves prognostic accuracy for multi-endpoint outcomes.

0 citationsRead paper

Recent publications

Latest Papers

FLEURS-Kobani: Extending the FLEURS Dataset for Northern Kurdish

Mar 31, 2026

This study addresses the scarcity of publicly available speech benchmarks for Northern Kurdish (Kurmanji, KMR), which has hindered progress in automatic speech recognition (ASR) and speech translation. To bridge this gap, the authors introduce FLEURS-Kobani, the first KMR-focused dataset derived from the FLEURS benchmark, comprising 5,162 validated utterances (18 hours and 24 minutes) from 31 native speakers, enabling research in ASR, end-to-end speech-to-text translation (S2TT), and speech-to-speech translation (S2ST). Leveraging the Whisper v3-large architecture with Common Voice pretraining and a two-stage fine-tuning strategy, the proposed system achieves a word error rate (WER) of 28.11% and character error rate (CER) of 9.84% on ASR, and a BLEU score of 8.68 for end-to-end KMR→EN S2TT, establishing a foundational benchmark for low-resource Kurdish speech processing.

0 citationsRead paper

SENS-ASR: Semantic Embedding injection in Neural-transducer for Streaming Automatic Speech Recognition

Feb 17, 2026

This work proposes a semantics-enhanced streaming automatic speech recognition (ASR) approach to address performance degradation caused by the absence of future contextual information in low-latency scenarios. By integrating a knowledge distillation–based sentence-embedding language model with a context-aware semantic module, the method extracts high-level semantic cues from historical frames and injects them into a neural Transformer architecture. This integration effectively mitigates the context deficiency inherent in chunk-based streaming processing. Experimental results on standard benchmarks demonstrate a significant reduction in word error rate, validating the efficacy and novelty of leveraging semantic guidance to enhance streaming transcription quality.

0 citationsRead paper

Toward More Reliable Artificial Intelligence: Reducing Hallucinations in Vision-Language Models

Dec 08, 2025

Visual language models (VLMs) suffer from hallucination—generating factually inconsistent or image-ungrounded statements. To address this, we propose a training-free, parameter-free self-correction framework that iteratively refines model outputs via uncertainty-guided visual re-attention. Specifically, it quantifies token-level uncertainty across four dimensions—token entropy, attention dispersion, semantic consistency, and claim confidence—to dynamically identify and crop unreliable image regions, followed by response revision. The method requires no fine-tuning, external data, or architectural modifications. Evaluated on Qwen2.5-VL-7B, it reduces hallucination rate by 9.8 percentage points and improves object existence accuracy by 4.7 percentage points, outperforming existing training-free baselines. Our core contribution lies in the tight coupling of multi-dimensional uncertainty modeling with dynamic visual attention, enabling interpretable, lightweight, and reliable VLM inference without parameter updates.

0 citationsRead paper

Sparse deepfake detection promotes better disentanglement

Oct 07, 2025

To address the trade-off between model interpretability and detection performance in deepfake speech detection, this paper introduces a Top-K sparse activation mechanism into the embedding layer of the AASIST architecture, proposing a latent-space sparse disentangled representation method. Inspired by sparse autoencoders, the approach enforces controllable sparsity to compel the model to focus exclusively on discriminative spoofing features, thereby achieving disentangled encoding of attack patterns in the latent space. To quantitatively evaluate disentanglement quality, we design integrity and modularity metrics based on mutual information. Experimental results on the ASVSpoof2021 LA and DF evaluation sets demonstrate an EER of 23.36% with 95% sparsity—significantly improving both detection accuracy and interpretability while preserving model lightweightness.

0 citationsRead paper

A dependent and censored first hitting-time model with compound Poisson processes

Apr 11, 2025

This paper addresses the challenge of modeling bivariate first-hitting times under dependence and right-censoring—common in clinical dual-endpoint settings such as concurrent liver and kidney injury. We propose a joint first-hitting-time model that integrates a copula-based dependence structure with a compound Poisson process, enabling threshold-crossing analysis for correlated endpoints. Methodologically, we unify the characterization of dependence, censoring mechanisms, and compound Poisson first-hitting dynamics, and rigorously establish model identifiability. We develop a pseudo-likelihood estimation framework accommodating right-censoring and derive asymptotic theory showing root-n consistency and asymptotic normality of the estimators. Monte Carlo simulations confirm robust finite-sample performance. Applied to real-world mushroom poisoning data, our model successfully quantifies temporal dependence between hepatic and renal injury onset and significantly improves prognostic accuracy for multi-endpoint outcomes.

0 citationsRead paper