Key Phrase Recognition for Low-Power FPGA Devices

MosChip designed and developed a custom Key Phrase Recognition solution for low-power FPGA devices, enabling always-on voice recognition within limited power, memory, and processing resources. The solution provides a reusable foundation for voice-enabled applications and simplifies AI deployment on resource-constrained hardware. The engagement covered five areas:

• Custom neural network architecture: A lightweight network designed around FPGA constraints such as limited DSP slices, memory bandwidth, and power budgets. It minimizes computational overhead while keeping high accuracy in detecting key phrases.
• Comprehensive model development: The full lifecycle, from dataset preparation (cleaning, augmenting, and labelling voice samples) through training, fine-tuning, and evaluation across multiple datasets, to a deployment-ready pipeline. Customers can move from prototyping on general-purpose platforms to running on FPGA hardware without major rework.
• FPGA deployment and demonstration: The trained model was quantized, compiled with the Lattice toolchain, and deployed on the Lattice iCE40 FPGA. In the working demo, voice command phrases such as “machine on” and “machine off” are recognized in real time to trigger corresponding actions, proving the approach for industrial, IoT, and consumer use.
• Pre-trained and documented reference model: A pre-trained model with step-by-step documentation, reference pipelines, and deployment instructions. Customers can reuse it directly, retrain it with their own data, or extend it to support additional voice commands.
• Validation and reporting: Functional testing and evaluation reports covering accuracy, latency, and FPGA resource utilization, giving customers confidence that the solution performs consistently across different FPGA environments.