Comprehensive research knowledge base: working papers, curated library, literature reviews, and research artifacts.
Research papers from leading scholars in our field
["Yohannes Ayana Ejigu","Tesfa Tegegne Asfaw","Surafel Amsalu Tadesse"] (2025)
This paper proposes a noise-robust end-to-end ASR framework for Amharic using CNNs, BiGRUs, and Connectionist Temporal Classification. Trained on 20,000 noisy Amharic utterances, the system achieves a 7% word error rate, demonstrating strong performance in real-world noisy environments and providing a practical model for low-resource ASR development.
["Manuela Cordeiro","Joao C. Ferreira"] (2025)
This paper investigates how blockchain, Decentralised Identifiers (DIDs), digital twins, and smart contracts can be combined to establish verifiable digital identities for agricultural products. It identifies major limitations in current traceability systems, proposes a layered architecture integrating digital twins and DIDs, and highlights gaps in interoperability, governance, and maturity. A cold-chain scenario is used to demonstrate the practical application of the proposed model.
Weiyao Kang, Bingjia Shao, Shan Du, Hongquan Chen, Yong Zhang (2024)
This study examines how social and technical attributes of voice assistants (VAs) drive consumer evaluation behaviour, such as continuance intention and positive word-of-mouth. Drawing on attachment theory and socio-technical systems theory, the authors propose that VA evaluations are shaped by two forms of attachment — emotional and functional — which arise from different VA attributes. Emotional attachment is influenced by interactivity, natural speech, and design aesthetics, while functional attachment is strengthened by accuracy, connectivity, and personalization. Using a two-wave longitudinal survey of 462 participants in China, analysed via PLS-SEM and fsQCA, the study confirms these relationships and identifies configuration paths leading to high continuance and word-of-mouth intentions. It also finds that social and technology anxiety moderate several attribute–attachment relationships. The findings provide new insights for improving both the theory and practical design of voice assistants.
Mohammad Zia Ur Rehman, Devraj Raghuvanshi, Nagendra Kumar (2024)
This paper presents KisanQRS, a deep learning-based automated query–response framework designed to support agricultural decision-making by rapidly retrieving relevant answers to farmers' queries. Leveraging a dataset of 34 million call logs from India's Kisan Call Centers (KCC), the system combines semantic and lexical similarity measures to cluster similar queries, employs an LSTM-based model for accurate query mapping, and introduces a novel answer retrieval module that clusters and ranks candidate responses. Experiments across multiple Indian states demonstrate that KisanQRS outperforms traditional approaches by achieving up to 97% accuracy in query mapping and an NDCG score of 96.2% in answer retrieval. The method provides a scalable and efficient pipeline for enabling rapid agricultural advisory services, particularly in contexts where manual call centre capacity is limited.
["Ivana Radic","Andrea Gardeazabal"] (2024)
This working paper explores how blockchain-enabled digital identities and data wallets can empower smallholder farmers by improving financial inclusion, traceability, and data governance. Using case studies from CIMMYT collaborations with Bluenumber and Identi, it demonstrates how secure digital identities enable farmers to control their data, access financial services, meet compliance requirements, and integrate into formal value chains.
Jonathan Steinke, Jerusha Onyango Achieng, James Hammond, Selamawit Sileshi Kebede, Dejene Kassahun Mengistu, Majuto Gaspar Mgimiloko, Jemal Nurhisen Mohammed, Joseph Musyoka, Stefan Sieber, Jeske van de Gevel, Mark van Wijk, Jacob van Etten (2019)
The paper examines whether household-specific agricultural advisory messages can be delivered through mobile phones using minimal data inputs. Across study sites in Ethiopia, Kenya, and Tanzania, the authors collected household indicators and farmers' ranked preferences for agricultural information. Using Bradley-Terry models, they show that 2–4 simple household variables—such as recent changes in agricultural input use, labor availability, age, and region—can produce useful individualized rankings of advisory messages. The targeted 'top 3' suggestions generated by the models were significantly better aligned with farmers' preferences than random message selections. The study concludes that rapid, low-burden data collection via mobile phones (e.g., IVR or USSD) can enable scalable, household-specific targeting of digital agricultural advice.
["Eshete Derb Emiru","Yaxing Li","Shengwu Xiong","Awet Fesseha"] (2019)
This conference paper presents a DNN-based acoustic modeling approach for Amharic ASR, using grapheme-to-phoneme conversion and syllable-based pronunciation dictionaries. The study evaluates several DNN architectures and demonstrates substantial improvements in word error rate for Amharic, a low-resource and morphologically rich language.