Hirofumi Inaguma
Postdoctoral Research Scientist at Meta
Menlo Park, California, United States
Summary
🤩
Rockstar🎓
Top SchoolHirofumi Inaguma is a postdoctoral research scientist at FAIR, Meta AI, specializing in speech-to-speech translation with nine years of experience across end-to-end ASR, speech translation, and language modeling. He earned his Ph.D. at Kyoto University and has a strong track record of research internships at Microsoft, Johns Hopkins, and IBM, contributing to streaming and multilingual ASR efforts. At Meta and in open-source, he has materially improved speech translation toolchains—contributing to flagship toolkits like fairseq and ESPnet by refactoring XMTransformer components, adding Conv2dSubsampler, and integrating R-Drop for more robust training. His work bridges deep research and production-grade engineering, often focusing on backend training pipelines and architecture-level enhancements. Based in Menlo Park, he combines academic rigor with practical system maintenance and cross-repository merges, a sign of fluency in collaborative, large-scale ML codebases. Colleagues describe him as someone who elevates model performance through meticulous code hygiene as much as algorithmic innovation.
9 years of coding experience
Bachelor's degree, Computer Science, Bachelor's degree, Computer Science at Kyoto University
English, Japanese
Github Skills (11)
pytorch10
machine-learning10
artificial-intelligence10
nlp10
python10
fairseq10
model-optimization9
neural-network9
convolutional-neural-networks9
deep-learning8
ctc7
Programming languages (2)
Github contributions (5)
espnet/espnet
Jun 2018 - Nov 2021
End-to-End Speech Processing Toolkit
Role in this project:
Back-end Developer
Back-end DeveloperContributions:32 reviews, 1106 commits, 114 PRs in 3 years 5 months
Contributions summary:Hirofumi primarily focused on modifying the training procedure and core functionalities of the speech translation system, specifically disabling a component of the CTC model, which indicates a focus on the system's backend and performance characteristics. Furthermore, the user was involved in integrating recent updates to the repository's codebase, merging branches to update and maintain its core functionality. These changes involved alterations to training scripts and model structures, which highlights a strong understanding of the system's internals.
speech-recognitionspeech-separationchainerspoken-language-understandingspeech-processing
facebookresearch/fairseq
Sep 2022 - Oct 2022
Facebook AI Research Sequence-to-Sequence Toolkit written in Python.
Role in this project:ML Engineer
ML EngineerContributions:22 reviews, 17 commits, 23 PRs in 1 month
Contributions summary:Hirofumi primarily focused on modifying and enhancing the speech translation modules within the fairseq library. Their contributions involved refactoring and reformatting code, specifically related to the XMTransformer model, including argument parser adjustments, convolution operations, and overall code cleanup. The user also implemented changes to incorporate Rdrop for improved model training and performance, demonstrating expertise in machine learning optimization techniques. Further contributions included the addition of Conv2dSubsampler and modifications to the multi-task learning framework, showcasing experience in model architecture and training pipeline enhancements.
pytorchnlpsequencepythontransformer-architecture
Find and Hire Top DevelopersWe’ve analyzed the programming source code of over 60 million software developers on GitHub and scored them by 50,000 skills. Sign-up on Prog,AI to search for software developers.