A paper authored by Amir Hussein, a PhD student in the Department of Electrical and Computer Engineering (ECE) and the Center for Language and Speech Processing (CLSP), as well as a Fellow in the Amazon Initiative for Interactive AI (AI2AI) program, received the Best Paper Award at the International Conference on Spoken Language Translation (IWSLT 2025). The event was held from July 31 to August, 2025, in Vienna, Austria, in conjunction with the Annual Meeting of the Association for Computational Linguistics (ACL), one of the world’s premier conferences in natural language processing. Second author Cihan Xiao, a PhD student in ECE and CLSP, played a key role in the work. Like Hussein, Xiao focuses on translating speech in one language into text in another, and has developed innovative ways to improve OpenAI’s Whisper model for this purpose.
Hussein and his colleagues’ award-winning paper, “HENT-SRT: Hierarchical Efficient Neural Transducer with Self-Distillation for Joint Speech Recognition and Translation,” tackles a major challenge in automatic speech translation: delivering fast, accurate translations with minimal delay and computational load. Their work introduces efficient neural architectures that enable high-quality translations suitable for real-time use on mobile devices and in interactive environments. Rather than replacing approaches like simultaneous interpretation, the team’s method reduces the time between hearing a full utterance and generating a translation, making the technology more practical and adaptable across diverse settings. Notably, the proposed joint ASR-ST modeling approach not only enhances speech translation performance but also achieves state-of-the-art ASR accuracy in multilingual settings, establishing a new benchmark.
The paper was co-authored by Matthew Wiesner, a researcher at the Human Language Technology Center of Excellence (HLTCOE); Dan Povey of Xiaomi; Leibny Paola Garcia, a research scientist at CLSP; and Sanjeev Khudanpur, an associate professor at ECE and CLSP.