Thanathai Lertpetchpun
tʰánátʰāj lɤ̂tpʰétpʰān
PhD Student in Electrical and Computer Engineering
University of Southern California
I research speech processing, focusing on speaker & emotion recognition, text to speech, voice conversion, accent conversion, and anti-spoofing. I am advised by Shrikanth Narayanan.
About Me
Hi, I am Thanathai and from Thailand (you can probably guess that from my first name). I am a second-year PhD student at the University of Southern California, where I'm a member of the Signal Analysis and Interpretation Laboratory (SAIL). My research interests are in speech processing, spanning both the analysis and generation of speech. Early on I worked on speaker verification and voice anti-spoofing, and I later built state-of-the-art speech emotion recognition systems that ranked 1st in emotional-attribute prediction and 2nd in categorical-emotion prediction at the Interspeech 2025 Naturalistic Conditions Challenge. These days I focus on text-to-speech, developing learning-free and controllable frameworks to control the accent strength of non-native speakers. I am also working on the privacy and security of speech systems, including membership-inference attacks and speaker-poisoning methods that erase voice-cloning capability in zero-shot TTS.
Before starting my PhD, I received my Bachelor's Degree in Computer Engineering from Chulalongkorn University, Thailand, where I also worked as a research assistant with Ekapol Chuangsuwanich on speaker verification and anti-spoofing. I've also had the opportunity to intern at Tencent Thailand in 2021, where I worked on news and song recommendation systems.
News
[Jul 2026] One paper was accepted to the Workshop on Responsibly Enabling Data for Foundation Models (ReData) at COLM 2026!
[Jul 2026] One paper was accepted to Interspeech 2026!
[Apr 2026] One paper was accepted to KDD 2026!
[Jan 2026] Three papers were accepted to ICASSP 2026
[Aug 2025] Three papers were accepted to Interspeech 2025!
[Aug 2023] One paper was accepted to Interspeech 2023!
Projects
ARTS: Anonymous Real-Time Speech
Oct 2024 - Present | USC - JHU - Meaning Company
- Developing learning-free framework for multilingual TTS on British, Spanish, and Indian Accent.
- Proposed Accent Vector, a controllable accent manipulation method for multilingual TTS that requires no accented training data.
- Proposed a speaker poisoning framework that erases a target speaker's voice-cloning capability in zero-shot TTS to protect against unauthorized voice generation. (1 COLM ReData Workshop 2026 publication)
- Analyzed the effect of speaker embedding on accent strength using phonemes rules to control the effect. (1 ICASSP2026 publication)
- Proposed frameworks to develop state-of-the-art speech emotion recognition (SER) for predicting both emotional attributes and categorical emotions. They achieved 1st rank in emotional attributes and 2nd rank in categorical attributes in Naturalistic Conditions Challenge (2 Interspeech2025 publications)
Speaker Recognition and Anti-spoofing System
Aug 2022 - May 2024 | Chulalongkorn University
- Proposed an approach to amplify artifacts in spoofed utterances using speech enhancement and show its robustness on different countermeasures and speech enhancement models (1 Interspeech2025 publication).
- Proposed a new normalization layer to address the problem of mismatched emotions and languages in speaker verification. (1 Interspeech2023 publication).
Awards & Honors
Student Grant
May 2026
Awarded a Student Travel Grant for ICASSP 2026.
SER Challenge
Aug 2025
Rank 1st in attribute prediction and 2nd in category prediction on the Speech Emotion Recognition in Naturalistic Conditions Challenge.
Student Grant
Aug 2023
Awarded a Student Travel Grant for Interspeech 2023.
Academic Service
Reviewer for the following venues:
- Interspeech 2026
- IEEE Spoken Language Technology Workshop (SLT) 2026
- COLM ReData Workshop 2026