TinyDenoiser: RNN-based Speech Enhancement on Multi-Core MCU with Mixed FP16-INT8 Post-Training Quantization

Learn about optimizing Speech Enhancement algorithms using Recurrent Neural Networks for microcontroller deployment in this technical talk from KU Leuven researcher Manuele Rusci. Explore a novel methodology that combines parallel computation of LSTM/GRU units with efficient memory management on multi-core RISC-V MCUs. Discover how a mixed FP16-INT8 post-training quantization approach enables up to 4x speedup while maintaining accuracy for TinyDenoiser models with up to 1.24M parameters. Examine the implementation details that achieve 10x better energy efficiency compared to existing single-core MCU solutions, making this approach particularly valuable for resource-constrained edge devices requiring speech enhancement capabilities.