Children and Low-Resource ASR
Improving recognition where data quality and language coverage are limited and high speaker diversity is present.
:: DR THOMAS ROLLAND ::
Bridging human and machine intelligence through robust speech AI.
I am a Postdoctoral Researcher at INESC-ID in Lisbon building speech systems that remain useful when data is scarce, noisy, domain-shifted, or unevenly distributed. My research focuses on advancing speech technology through the design of parameter-efficient architectures, large-scale synthetic data augmentation, and post-training strategies aimed at improving robustness, adaptability, and fairness across diverse languages, speakers, and domains.
Affiliation
INESC-ID, Lisbon
Focus
ASR, TTS, Efficient ML, SpeechLMs, PEFT
Current lens
Efficient ML, Children's speech and Post-Training
> RESEARCH_AGENDA
Research that stays useful when the data gets messy.
I focus on the parts of speech AI that are easiest to ignore and hardest to fake: data scarcity, domain mismatch, evaluation quality, and adaptation cost.
Across children's speech, low-resource ASR, Post-training, and synthetic data, the through-line is the same: make systems more robust, more comparable, and more deployable in real research and product contexts.
> TRACKS_03
Parameter-Efficient Learning
Knowledge distillation, Quantization, Shared-weights to adapt models with lower compute power.
Synthetic Data generation for SFT
Generation and benchmarking work that makes synthetic speech post-training trustable, and deployable.
> EXPLORE_PATHWAYS
The homepage stays selective. The detailed record lives on dedicated pages.
> CONTACT_NODE