Deep Learning Speech Commands Project

End-to-end keyword spotting project built on the Google Speech Commands dataset. The project supports:

• training a baseline KWS model
• evaluation on the test split
• single-file prediction
• TensorFlow Lite export for Float32, Int8, and experimental Int4 variants
• robustness evaluation under local noise samples at multiple SNR levels

The 12 output classes are: yes, no, up, down, left, right, on, off, stop, go, unknown, silence.

Project Files

• train.py: trains the baseline Keras model and saves metrics/plots.
• evaluate.py: evaluates the saved Keras model on the clean test split.
• predict.py: predicts the class for one WAV file.
• quantize.py: exports the trained model to Float32, Int8, and experimental Int4 TFLite formats.
• evaluate_robustness.py: evaluates TFLite models on clean audio and noisy audio at multiple SNR levels.
• src/config.py: paths and model/audio constants.
• src/dataset.py: dataset extraction, manifests, TensorFlow datasets, and log-mel preprocessing.
• src/model.py: CNN model definition.
• src/noise.py: local noise loading and SNR-based audio mixing helpers.

Dataset Requirements

You need:

1. The Google Speech Commands archive.
2. Optional local noise files for robustness testing.

Default archive path expected by the project: D:\...\speech_commands_v0.02.tar.gz

Local noise folder expected by the robustness script: D:\...\data\noise_samples

Example noise files already supported:

• Babble_1.wav
• Cafe_1.wav
• Traffic_1.wav
• AirConditioner_1.wav

Create a "data" subfolder and extract it.

Step-by-Step: Run Locally in VS Code Terminal

Open the project folder in VS Code and make sure the terminal is already inside: D:\CVprojects\DL-Project

Step 1: Install dependencies

pip install -r requirements.txt

Step 2: Train the baseline model

Quick verification run:

python train.py --epochs 3 --limit-train 6000 --limit-val 1200 --limit-test 1200

Full training run:

python train.py --epochs 12

This saves:

• artifacts/models/best_model.keras
• artifacts/plots/training_history.png
• artifacts/plots/confusion_matrix.png
• artifacts/reports/metrics.json
• artifacts/reports/classification_report.txt

Step 3: Evaluate the baseline Keras model

python evaluate.py

Step 4: Run single-file prediction

python predict.py --wav "D:\CVprojects\DL-Project\data\yes\004ae714_nohash_0.wav" --plot

Step 5: Export TensorFlow Lite models

python quantize.py

This writes:

• artifacts/models/model_float32.tflite
• artifacts/models/model_int8.tflite
• artifacts/models/model_int4.tflite
• artifacts/models/quantization_summary.json

Note: the Int4 path is experimental and depends on TensorFlow Lite support in your local build.

Step 6: Run robustness evaluation

Quick demo run:

python evaluate_robustness.py --limit-test 100

Full robustness run:

python evaluate_robustness.py

This writes:

• artifacts/reports/robustness_results.csv
• artifacts/reports/robustness_accuracy_table.csv
• artifacts/reports/deployment_summary_table.csv
• artifacts/reports/robustness_summary.json

Best Demo Order for Instructor

If you want to demonstrate the full project quickly, run these in order:

python quantize.py

python evaluate_robustness.py --limit-test 100

If your instructor wants baseline training too, run this before the above two commands:

python train.py --epochs 3 --limit-train 6000 --limit-val 1200 --limit-test 1200

Important Output Files to Show

• artifacts/models/best_model.keras
• artifacts/models/model_float32.tflite
• artifacts/models/model_int8.tflite
• artifacts/models/model_int4.tflite
• artifacts/plots/training_history.png
• artifacts/plots/confusion_matrix.png
• artifacts/reports/metrics.json
• artifacts/reports/classification_report.txt
• artifacts/reports/robustness_accuracy_table.csv
• artifacts/reports/deployment_summary_table.csv

Notes

• Audio is processed as 16 kHz mono and padded or trimmed to 1 second.
• Features are normalized log-mel spectrograms.
• The baseline model is a CNN trained with TensorFlow/Keras.
• The robustness script evaluates clean audio and noisy audio at 20 dB, 10 dB, 0 dB, and -5 dB SNR.
• The TFLite Int4 path should be treated as experimental unless separately validated in your exact TensorFlow build.