Speech Enhancement & Dereverberation

At Logitech I improved the deep learning speech enhancement model running on thousands of video conferencing devices worldwide, including the Rally Bar range, raising audio quality across real-world rooms, noise conditions and languages. The models ran in real time on low-power embedded chips, processing a new audio frame every 10 ms with compute budgets as low as 2 MFLOPS.

I led the end-to-end development of a deep learning dereverberation system that improved PESQ by around 1.0, from model architecture and training through hyperparameter optimisation to production deployment, and prototyped early Transformer and convolutional architectures that shaped later production models.

I also designed and ran large-scale audio dataset collection and annotation across rooms, devices, languages and noise conditions, and built the evaluation framework and objective metrics used to benchmark every speech enhancement and dereverberation model before release. I contributed to the early development and prototyping of Logitech Sight, later recognised as one of TIME's Best Inventions of 2024.