Side-end speech recognition model compression method based on knowledge distillation
By combining knowledge distillation techniques from deep neural networks and Hamiltonian neural networks, the student model structure is dynamically adjusted, solving the adaptability problem of the side-end speech recognition model on resource-constrained devices and achieving efficient and stable speech recognition results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Filing Date
- 2026-04-29
- Publication Date
- 2026-07-17
AI Technical Summary
Existing side-end speech recognition models are difficult to compress effectively on resource-constrained devices, and existing knowledge distillation methods fail to fully utilize the intermediate layer features of the teacher model, resulting in insufficient modeling of temporal evolution relationships in the student model and poor adaptability.
This paper adopts a method combining deep neural networks and Hamiltonian neural networks. Through knowledge distillation, intermediate layer features are extracted from the teacher model and passed to the student model. The structural parameters of the student model are dynamically adjusted, and the model is optimized in combination with edge device resource conditions. The dynamic modeling of Hamiltonian neural network is introduced to optimize temporal feature learning.
It improves speech recognition accuracy, reduces computational complexity and storage requirements, enhances the model's adaptability and operating efficiency on edge devices, and achieves stable, low-latency speech recognition.
Smart Images
Figure CN122416992A_ABST