Ryo Magoshi
M.S. student, Kyoto University, Kyoto, Japan
About
I am a master's student at Kyoto University working on speech processing. My research centers on multilingual speech recognition, articulatory features, and SpeechLLM.
Research Interests
- Speech Recognition: Multilinguality, Low-Resource and Zero-Shot Languages, Articulatory Features
- Post-Training of LLMs and Speech LLMs: Domain Adaptation, Reinforcement Learning (RLVR, Preference Optimization), Multilingual Capabilities
- Spoken Dialogue Systems: Cascaded or Full-duplex Dialogue Systems
Publications
- [C1]Proc. IEEE Spoken Language Technology Workshop (SLT), 2026
- [C2]Proc. INTERSPEECH, 2026Nominated for Best Student Paper Award
- [C3]
- [C4]Proc. INTERSPEECH, 2026
- [C5]Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2026
- [C6]
Projects
- glideTRL-based post-training library for LLMs, Speech LLMs, and Vision LLMs. Supports SFT and RL (GRPO, GSPO), autoregressive-decoding evaluation (WER / CER / BLEU / ROUGE), composable YAML configs, and plugins for custom encoders, projectors, rewards, and metrics.
- Megatron-SpeechSpeech LLM training framework built on Megatron Core, with tensor parallelism and Hugging Face checkpoint conversion.
- mightningMinimal Lightning-like PyTorch wrapper for single-node multi-GPU training, built on pure DistributedDataParallel, with W&B logging, metric-based checkpointing, and 16-bit mixed precision.