Kaidtarek/Voice-Assisted-TC-ResNet-model-with-benchmark-

TC-ResNet keyword spotting for Algerian Darja smart-home commands on ESP32 (Heltec LoRa32 V2) — Edge Impulse expert-mode training, EON-compiled INT8 deployment, 94.10% validation accuracy. # TC-ResNet Keyword Spotting Pipeline **Figure 1. TC-ResNet architecture — Conv1D stem with 4 temporal residual blocks and a 17-class softmax output** Welcome to the TC-ResNet module for the **Keyword Spotting for Smart Home Control in Algerian Darja Using TinyML** project. This folder contains the training code, Edge Impulse project configuration, and the deployment benchmark for the TC-ResNet model on the ESP32 (Heltec WiFi LoRa 32 V2) board, using the EON Compiler for compiled inference. > **⭐ If you find this work useful, please consider giving it a star!** ## 📂 Project Structure | Folder / File | Description | |---|---| | `train.py` | Edge Impulse expert-mode Keras training script | | `EI-TC-ResNet-benchmark/` | PlatformIO firmware for on-device inference using the EON-compiled model — **fully tested** | | `summary/` | Edge Impulse screenshots, confusion matrix, accuracy charts, and MFCC settings | ## 🧠 Model Architecture TC-ResNet is a 1D temporal convolutional residual network, trained in Edge Impulse expert mode and deployed via the EON Compiler, which generates optimized standalone C++ inference code — bypassing runtime interpreter overhead. | Layer | Details | |---|---| | Input | Reshape to (frames, 13 MFCC features) | | Stem | Conv1D, 16 filters, BN, ReLU | | Residual block 1 | 24 filters, kernel 9, dilation 1 | | Residual block 2 | 24 filters, kernel 9, dilation 2 | | Residual block 3 | 32 filters, kernel 9, dilation 4 | | Residual block 4 | 32 filters, kernel 9, dilation 8 | | Head | GlobalAveragePooling1D → Dropout(0.40) → Dense(17, softmax) | ## ⚙️ Edge Impulse & Training Configuration **Figure 2. Edge Impulse impulse design — 2-second time-series audio window, MFCC block, and Classifier** **Figure 3. MFCC feature extraction settings — 13 coefficients, 20ms frame length, 10ms stride, 32 Mel filters, FFT 512** | Parameter | Value | |---|---| | Dataset | 4,800 Algerian Darja commands + 300 noise files | | Classes | 17 (16 commands + 1 …