Kaidtarek/Voice-Assisted-DS-CNN-model-with-benchmark

DS-CNN keyword spotting for Algerian Darja smart-home commands on ESP32 (Heltec LoRa32 V2) — TensorFlow/Keras training, INT8 quantization, and TFLM/MicroTVM deployment. # DS-CNN Keyword Spotting Pipeline **Figure 1. TFLM deployment setup and serial monitor output during DS-CNN inference** Welcome to the DS-CNN module for the **Keyword Spotting for Smart Home Control in Algerian Darja Using TinyML** project. This folder implements the full pipeline for the DS-CNN model: training with TensorFlow/Keras, INT8 quantization, TFLite model export, and deployment on the ESP32 (Heltec WiFi LoRa 32 V2) using TensorFlow Lite for Microcontrollers (TFLM). > **⭐ If you find this work useful, please consider giving it a star!** ## 📂 Project Structure | Folder / File | Description | |---|---| | `train_ds_cnn.py` | Dataset loading, MFCC extraction, model training, INT8 export, and C-array generation | | `export_model.py` | Generates MicroTVM AOT artifacts from the trained model | | `TFLM-benchmark/` | PlatformIO firmware for running the model on-device using TFLM — **fully tested** | | `MicroTVM-benchmark/` | PlatformIO firmware for running the model using Apache MicroTVM — provided for experimentation | | `models/` | Output directory for trained model artifacts (`DS-CNN.h5`, `DS-CNN.tflite`, `DS-CNN.cc`) | | `summary/` | Training output log and setup images | ## 🧠 Model Architecture The V2 DS-CNN is a 2D depthwise separable CNN designed to fit within the ESP32 memory constraints, inspired by the Hello Edge baseline. | Layer | Details | |---|---| | Input | (199, 13, 1) — MFCC features | | Stem | Conv2D, 12 filters, BN, ReLU | | DS Block 1 | Depthwise-separable, 12 filters, stride (2,1) | | DS Block 2 | Depthwise-separable, 12 filters, stride (1,1) | | DS Block 3 | Depthwise-separable, 16 filters, stride (2,1) | | DS Block 4 | Depthwise-separable, 20 filters, stride (1,1) | | Head | GlobalAveragePooling2D → Dropout(0.25) → Dense(17, softmax) | ## ⚙️ Training Configuration | Parameter | Value | |---|---| | Dataset | 4,800 Algerian Darja commands + 300 noise files | | Classes | 17 (16 commands + 1 noise) | | Audio Input | 16 kHz mono, 2-se …