Huck Yang

About

I am interested text-space alignment of speech, voice-agents, and multimodality. I worked at NVIDIA Research and Amazon ASR-LM, working with Andreas Stolcke, and as a research intern at Google (now DeepMind Audio), co-hosted by Bo Li and Yu Zhang in Tara N. Sainath's team.

🎓 My Ph.D. topic is on noise-robust voice model adaptation (now post-training), advised by Prof. Chin-Hui Lee.

🧬 I visited Prof. Jesper Tegnér's group on self-evolutionary algorithms and interned at TSMC in mixed-signal IC design before starting my PhD.

• Post-Training: the first prompt-adaptation method (i.e., trainable inputs plus label mappings) to frozen acoustic models [ICML 21], concurrent with prefix-tuning (ACL 21). Best paper nominee 🎖️ on a multilingual study in [Interspeech 23]; a Google affiliated patent [ICASSP 23].
• Generative Error Correction: the first series of n-best hypotheses based generative error correction (GER) in ASR & translation post-training in [ASRU 23] and co-invented Whispering-LLaMA [EMNLP 23]; HyPoradise in [NeurIPS 23], and speech post-training to LLM [ICLR 24]. Best industry paper honoronable mentioned award 🎖️ on multimodal n-best correction in [ACL 25].

⚛️ Fun fact: I also work on Quantum ML part-time for fun, where I created the first variational circuit based speech [ICASSP 21] and language understanding [ICASSP 22] and received the Xanadu AI Quantum ML Award in 2019; recently, on quantum parameter adaptation for LLMs in [ICLR 25].

Recent Tutorials