Research Scientist, Shanda AI Research Tokyo
I work on speech and language AI, after a Ph.D. in Informatics at Kyoto University. My research covers robust speech processing, multi-talker ASR, and speech-language modeling.
The problem I keep returning to is how speech models can use strong learned priors while staying faithful to what was actually said: recovering speech from noise, attributing it to the right talker, and preserving pronunciation in Japanese ASR and TTS.
Robust Speech Processing
Recover reliable speech evidence from noisy, reverberant, and otherwise degraded observations.
Multi-Talker & Speaker-Grounded Speech
Understand what was said, who said it, and which acoustic evidence belongs to which source in overlapping speech scenes.
Speech-Language Modeling
Use linguistic knowledge in speech and language models while preserving pronunciation, spoken-language properties, and acoustic fidelity.
Reliable Perception
IEEE/ACM TASLP, 2024
Passing spectrogram encoding to both the enhancement front-end and the ASR back-end consistently improved recognition on the real CHiME-4 evaluation sets, outperforming methods including DEMUCS and Conv-TasNet.
Evidence Attribution
IEEE TASLP, 2026
Letting LLM layers query serialized acoustic memory throughout generation gave consistent improvements over static-prefix prompting on LibriMix.
Speech–Language Alignment & Grounding
arXiv, 2026 · Preprint
Binding each written span to its realized reading in the recognition target improved lexical-reading recovery across five Japanese benchmarks without sacrificing readable transcription.
