Hybrid ASR for constrained robots
HMM and deep learning fusion for speech recognition on embedded hardware.
Speech recognition on a robot has a hard budget: no cloud round trip, a few watts, and a deadline measured in milliseconds. Fusing a classical HMM front end with a deep acoustic model hit up to 26.59x faster inference on a Jetson Xavier NX than the neural baseline alone. Published at IEEE MoSICom 2024 in Dubai.
All projects
- Meta-ROS
Systems and robotics
- MicroRWKV
Model architecture
- Automated stock prediction pipeline
MLOps