This is simple weekend project to recognize and record sounds in real time with tensorflow an keras. It has 3 python entrypoints:

• real_time_classifier.py: It constantly records 2 second audio clips and runs a RNN against it and tries to classify whether the trained sound is there or not. If the confidence is greather or equal than 0.5 it saves a wav file in training-data/recognized/ folder.
• train.py: It picks all recordings from training-data/positive and training-data/negative, trains a basic RNN and saves the output in model.tflite
• web.py: It starts a flask app with some useful tools to collect training data and analyze the real time recordings. It has the following paths:
  • /recorder: The server in background constantly records sounds withing a 30 second window and this page contains a submit button to save a wav file with the recorder buffer into training-data/raw/. Later, those files can be trimmed and used as training data.
  • /cropper: It shows a page with all the raw recordings and allows to extract 2 seconds clips and mark them as positive sounds
  • /analyze: it shows a page with all the recognized sounds by the real_time_classifier and allows to mark them as positive or negative which can be later used for training
  • /report: It shows a simple json with the count of positive sounds by date

Build docker build -t classifier .

Run: docker run --rm -d --device /dev/snd -v "$(pwd)":/app classifier python real_time_classifier.py

• https://towardsdatascience.com/how-to-reduce-training-time-for-a-deep-learning-model-using-tf-data-43e1989d2961

• https://stackoverflow.com/questions/54431168/how-to-cache-layer-activations-in-keras

• https://github.com/kongkip/spela

• https://medium.com/swlh/how-to-run-gpu-accelerated-signal-processing-in-tensorflow-13e1633f4bfb

• https://towardsdatascience.com/how-to-easily-process-audio-on-your-gpu-with-tensorflow-2d9d91360f06