OLA: Office Laughter Analyzer is an audio-based real-time office laughter detection system.
main.py does training on AudioSet's feature data, and test the f1 score with evaluation data each epoch.
python main.py -m "model file name(without .py)" [--gpu] [--roll]
Example: python main.py -m dense --gpu --roll
-m / --model: Model to use (filename in directory ./Models/
--gpu: Use gpu to training.
--roll: Use Embedding Rolling to augment input data. The data will be augmented by 10 times.
Example: python main.py -m dense --noise 5 --trueWeight 0.2 --positiveWeight 1 --batch_size 256 --epoch 10
--noise N: Use noise augmentation. The data will be augmented by N times. Proven to be useless. Don't use.
--trueWeight W: Use weighted training, and set true labeled case weight to W. Note that weight value '1.0' means ~100 times stronger gradient compared to True Negative cases.
--positiveWeight W: Use weighted training, and set positive output case weight to W. Note that weight value '1.0' means ~100 times stronger gradient compared to True Negative cases.
--batch_size B: Set the batch size B.
--epoch E: Set the epoch E.
Example: python main.py -m lstm --eval
--eval: Activate evaluation mode, only conducts evauation based on saved model in ./TrainedModels.
live.py records audio with local microphone device, and saves the .wav file, and extract embedding vector using VGGish model downloaded, and do laughter inference with specified model.
python live.py -m "model file name(without .py)" --wavOutDir FILE_DIR --count N
Example: python live.py -m dense --wavOutDir recordedWavs --count 10
-m / --model: Model to use (filename in directory ./Models/
--wavOutDir PATH: Where to save recorded audio file (.wav). The audio files will be saved in ./Dataset/RawData/[PATH]
--count N: How many files to record? Note that one audio file has 10 seconds long.