USC SAIL · NSTC Fellow
周惶振
Speech / multimodal researcher for voice assistants— speech emotion recognition, subjective evaluation, and systems that hear affect as carefully as words.
I am a Postdoctoral Scholar–NSTC Fellow at the Signal Analysis and Interpretation Laboratory (SAIL) at the University of Southern California, working with Prof. Shrikanth S. Narayanan. I received my Ph.D. in Electrical Engineering from National Tsing Hua University, advised by Prof. Chi-Chun Lee.
My research centers on three threads: (1) speech emotion recognition under subjectivity and ambiguity— multi-label learning, annotator disagreement, calibration, fairness, and open evaluation via EMO-SUPERB; (2) conversational social signal processing—deception and belongingness in dialogue / small groups, and dyadic clinical affect such as depression-related speech; and (3) speech systems for assistants and multilingual / low-resource settings—unified ASR + SER, speech LLMs, expressive TTS evaluation, edge SER / KWS, and collaborations such as TaigiSpeech.
I have worked on speech and ML systems from research through deployment at Amazon Alexa, ITRI, and RealTek. I am motivated by building intelligent, expressive agents that integrate speech, language, and emotion for natural interaction.
Grouped by research theme. Selected papers under each; full list on Scholar and in the CV.
Theme
Modeling annotator disagreement, minority views, soft labels, and calibration so SER systems respect emotion ambiguity rather than collapsing to a single “ground truth.”
Theme
Affect and social meaning in dialogue: deception and perceived lying in interactive games, belongingness / satisfaction in small-group conversation, multimodal group emotion, and dyadic depression-related speech (including mentored work).
Theme
Subgroup disparity, debiasing without fragile demographic shortcuts, and confidence-oriented augmentation for fairer multi-label emotion recognition.
Open evaluation
Reproducible evaluation contracts for classical and speech-LLM SER: splits, metrics, prompts, parsing, and diagnostics so results stay auditable.
Theme
Unified ASR + multi-label SER for voice-assistant workloads, emotional-speech ASR analysis, stimulus-modality effects, and on-device SER / KWS under compute budgets.
Theme
Large audio / speech language models, expressive TTS evaluation, codec emotion preservation, and low-resource speech resources such as TaigiSpeech.
USC SAIL · SPAN
Contribution to SAIL’s Speech Production and Articulation Knowledge (SPAN) effort on real-time MRI of dynamic vocal-tract shaping. First-author work evaluates modern speech enhancement on rtMRI speech—objective quality plus downstream tasks (ASR, speaker, SER, demographics)— and releases practical guidance with an open resource so SPAN corpora are more usable beyond speech science. Manuscript in preparation (JASA-style venue); PDF will be posted after preprint or acceptance.
Full list on Google Scholar and in the CV appendix. Name in bold indicates Huang-Cheng Chou.
Additional scholarships and reviewing service are listed in the CV.
Open to conversations on speech AI, voice assistants, and research collaboration.