To promote academic exchange and advance research in linguistics-related fields, on May 15, 2026, Professor Yuan Jiahong, Associate Director of the Center for Interdisciplinary Studies of Language and Professor at the School of Humanities and Social Sciences, University of Science and Technology of China, visited our school to deliver a special lecture titled “From Speech Analysis to Model–Brain Alignment: A Case Study of Tone Recognition.” The lecture was hosted by Professor Liang Dandan, leader of the linguistics discipline at our school, and was attended by Professor Hao Ying, a Jiangsu Province specially-appointed professor, as well as young faculty members, Master's and Ph.D. students specializing in linguistics.

At the beginning of the lecture, Professor Liang Dandan expressed her gratitude to Professor Yuan Jiahong for his visit. She introduced that Professor Yuan has held positions at prestigious institutions and industry-leading companies, including the Department of Linguistics and the Linguistic Data Consortium at the University of Pennsylvania, the AI Lab of Liulishuo, and Baidu Research USA. He has also led numerous research projects funded by the U.S. National Science Foundation, the UK Economic and Social Research Council, and the National Social Science Fund of China, and has made significant contributions to fields such as speech recognition and language and health.

Professor Yuan Jiahong presented the background of the lecture, emphasizing the role of advances in artificial intelligence in driving progress in linguistics and neuroscience. He then systematically shared the latest research findings from his team. First, he focused on speech recognition and speech analysis based on the Transformer architecture and pre-training–fine-tuning methods. He demonstrated the powerful capabilities of Transformer-based models, represented by wav2vec2, in speech analysis. Through pre-training and fine-tuning, these models not only significantly improve speech recognition accuracy but also extract vowel and tone representations that are stronger than traditional formants and fundamental frequency. Moreover, t-SNE dimensionality reduction revealed that wav2vec2 representations are more robust in distinguishing vowels and tones, providing high-quality analytical tools for speech science research and applications.

The second part of the lecture focused on the cutting-edge direction of “model–brain alignment”. Professor Yuan systematically compared the processing mechanisms of speech models and the human brain in tone recognition. He noted that models exhibit categorization abilities even under “poverty of the stimulus” conditions—such as sparse or incorrect labels—resembling features of child language acquisition. Additionally, through multi-task learning, models can achieve speaker normalization without suppressing irrelevant information, offering a new perspective for studying the brain's speech perception mechanisms. These findings provide novel insights into understanding the shared principles between artificial intelligence and biological intelligence.

During the Q&A session, Shi Xinyuan, a faculty member in linguistics, and Hu Han, a Ph.D. candidate, raised questions regarding example-based learning and specific algorithms for speech recognition. Professor Yuan affirmed the academic perspectives of the young researchers and responded to each question in detail.

At the conclusion of the lecture, Professor Liang Dandan emphasized the importance of this lecture in broadening the academic horizons of researchers. She encouraged everyone to continue paying attention to core and emerging issues in linguistics and to explore and study them from the perspective of artificial intelligence. The lecture ended with renewed applause and appreciation from the audience for Professor Yuan's insightful presentation.