01
Amazon Nova 2
Audio and music generation, encoder, and codec contributions to Amazon's multimodal reasoning and generation models.
Senior Applied Scientist Amazon AGI
Byeonggeun Kim
I build models that understand and generate audio for natural, real-time conversations.
My research has progressed from efficient on-device and few-shot audio learning to audio encoders, neural codecs, and generation for Amazon AGI projects, including Amazon Nova. I currently focus on full-duplex speech-to-speech modeling.
Selected work
From on-device audio to generative systems.
Representative outcomes across efficient architecture, robust learning, and audio foundation models.
02
Continuous-valued Audio Language Modeling
Generative audio language modeling with masked next-token prediction and continuous-valued tokens.
03
BC-ResNets
Broadcasted residual learning for accurate, low-compute keyword spotting on resource-constrained devices.
04
DCASE 2021 Challenge Winner
A low-complexity, device-robust acoustic scene classification system combining efficient design and domain generalization.
Recent
News and milestones.
All publications-
Invited talk at Yale University (CPSC 7760): Efficient and High Fidelity Text-to-Audio Generation.
-
Contributed to the launch of Amazon Nova 2.
-
Invited talk at the Amazon Media & Entertainment ML+AI Summit.
-
Two papers accepted at ICML 2025, including IMPACT and continuous-valued audio language modeling.