Research & Publications

My research connects phonetics and computational linguistics, with a focus on speech variability, spoken-word recognition, and the reliability of speech-language models. I combine acoustic and perceptual studies with computational analyses of speech representations.

Research areas

Speech production and perception

I investigate how words change in spontaneous speech and how listeners recover words from reduced forms. My work on the English word just examines how neighboring sounds—particularly sibilants—and lexical predictability are associated with reduction. An ongoing perception study asks how listeners use sibilant and vowel duration across phonetic contexts.

Perception study (accepted poster at the 191st ASA Meeting (2026))

Spontaneous-speech ASR and speech representations

I study how computational models represent acoustic and phonetic variability. In a parameter-matched study using frozen HuBERT, adding a trainable fusion pathway improved recognition of spontaneous speech, but learned prosodic features did not yield a significant incremental gain over a capacity-matched control. Related work examines where phonetic variation becomes accessible across model layers.

Moon & Warner, submitted to ICPhS 2027

Uncertainty and speech reasoning

My ongoing research investigates whether intermediate representations in speech-reasoning models distinguish answerable and unanswerable inputs. I study lightweight monitoring mechanisms that prompt models to reassess potentially unsupported answers during generation. The goal is to understand when a model has enough evidence to answer and how it can respond when the available audio is insufficient.

Status: Ongoing research; no publication is claimed here.

Publications and manuscripts

Published

Moon, Ki Woong, and Natasha Warner. (2025). “Realization of just: Speech reduction in a high-frequency word.” Proceedings of Meetings on Acoustics, 56(1), 060005. DOI: 10.1121/2.0002080.

Submitted

Moon, Ki Woong, and Daniel Brenner. “Does a prosody-trained representation help beyond trainable fusion? A parameter-matched study with frozen HuBERT.” Submitted to ICASSP 2027. arXiv preprint.
Moon, Ki Woong, and Natasha Warner. “Layer-wise accessibility of phonetic variation in spontaneous just.” Submitted to ICPhS 2027.

In preparation

Moon, Ki Woong, and Natasha Warner. “Just reduced: Gradient acoustic reduction in a high-frequency function word.” Manuscript in preparation.
Warner, Natasha, Ki Woong Moon, Seongjin Park, and James M. McQueen. “Metrical segmentation and competition in native and cross-dialectal spoken word recognition.” Manuscript in preparation.

Conference presentations

2026 — Accepted poster. Moon & Warner, “Context-dependent use of sibilant and vowel duration in the perception of reduced just.” 191st Meeting of the Acoustical Society of America, Baltimore, Maryland.
2026 — Accepted poster. Moon & Brenner, “Prosody or adaptation? Disentangling gains in frozen self-supervised learning for automatic speech recognition.” 191st Meeting of the Acoustical Society of America, Baltimore, Maryland.
2025 — Invited talk. Moon & Brenner, “Enhancing HuBERT for spontaneous speech with acoustic feature fusion.” 6th Joint Meeting of the Acoustical Society of America and Acoustical Society of Japan, Honolulu, Hawaii.
2025 — Poster. Moon & Warner, “Realization of just: Speech reduction in a high-frequency word.” 188th Meeting of the Acoustical Society of America, New Orleans, Louisiana.
2024 — Poster. Moon, “Preceding word information for predicting speech errors in EFL speech.” LabPhon 19, Seoul, South Korea.
2024 — Poster. Moon, “Preceding word information for predicting speech errors in EFL speech.” 186th Meeting of the Acoustical Society of America, Ottawa, Canada.
2024 — Poster. Warner, Moon, Park, McQueen, & Albusairi, “Metrical segmentation across dialects.” 186th Meeting of the Acoustical Society of America, Ottawa, Canada.

Open research data

As a graduate research assistant in the Douglass Phonetics Lab, I curated previously collected Dutch and English speech-perception data into two publicly released datasets comprising approximately one million human phonetic-identification responses.

Dutch speech-perception dataset: DOI: 10.25422/azu.data.21498258.v1
English speech-perception dataset: DOI: 10.25422/azu.data.21498210.v1

The datasets are hosted by the University of Arizona Research Data Repository. This work involved organizing previously collected responses, stimuli, and supporting documentation for public release.

Ki Woong Moon