About

Audio representations and generative models, from problem formulation to evaluation.

I am Byeonggeun Kim (김병근), a Senior Applied Scientist at Amazon AGI. My current work is on fully causal neural codecs for full-duplex speech-to-speech; before that, I led audio encoder pretraining and audio/music generation work at Amazon. Earlier, at Qualcomm AI Research, I worked on efficient architectures, few-shot and open-set recognition, transfer learning, and domain adaptation.

Experience

Research in production and at scale.

Amazon AGI

Seattle, WA

Senior Applied Scientist

Audio and Speech AGI

  • Full-duplex speech-to-speech: lead codec research for full-duplex models, designing fully causal representations for both understanding and generation. Successive designs are adopted in internal speech-to-speech models.

Applied Scientist II

Toronto, ON & Seattle, WA

  • Audio representations: single-stage self-supervised encoder pretraining with curriculum masking, scaled to 400K hours (ICASSP 2025).
  • Generative audio/music: RVQ-based autoregressive audio/music generator for Amazon Nova 2 research; proposed masked next-token prediction for streaming continuous-valued generation (ICML 2025).
  • On-device acoustic event detection: developed and deployed compact models at Alexa AI, expanding the range of acoustic events supported on device (ICASSP 2024).

Qualcomm AI Research

Seoul, South Korea

Staff AI Researcher

Senior AI Researcher

AI Researcher

  • BC-ResNet: broadcasted residual learning for efficient keyword spotting (INTERSPEECH 2021).
  • DCASE 2021 Task 1A, 1st place: Residual Normalization for device-robust acoustic scene classification (papers).
  • Data-efficient learning: few-shot and open-set keyword spotting, transfer learning, and test-time adaptation.

Recognition & service

Awards, talks, and service.

  1. Invited talk · Yale University, CPSC 7760 Seminar

    Efficient and High Fidelity Text-to-Audio Generation: From Parallel Decoding to Causal Streaming.

  2. Invited talk · Amazon Media & Entertainment ML+AI Summit

    Autoregressive text-to-audio generation approaches.

  3. Alexa Perceptual Technology Award

    Q3/Q4 2023.

    View award photo

  4. DCASE 2021 Challenge · 1st Place (Task 1A)

    Residual Normalization for device-imbalanced acoustic scene classification.

    Official results Related papers

  5. Service

    Reviewer for ICASSP, INTERSPEECH, CVPR, ICCV, ECCV, NeurIPS, ACL, EMNLP, AAAI, WACV, and BMVC. Area Chair, ICASSP 2026; Session Chair, INTERSPEECH 2022.

Education

KAIST.

M.S., Electrical and Electronic Engineering

KAIST · Advisor: Soo-Young Lee

B.S., Electrical and Electronic Engineering & Business Administration

KAIST · Double major