This is the code accompanying blog posts (Here and Here) for training and running a laugh detector. This project is based on AudioSet and includes a rebalanced subset of their data and utilizes their pre-trained audio feature vectorizer vggish. I also followed the lead of a similar project for processing the AudioSet data and choosing network architectures
This repo contains code for running live inference of the presence of laughter, a few pre-trained models, code that was used for training, a subset of the AudioSet data containing a balanced set of laughter and non-laughter human speech and a dashboard for visualizing the amount of laughter.
Running live_inference.py will use your computer's microphone to capture audio and infer the presence of laughter in segmented chunks. These scores can be written to a time-stamped .csv and/or used to control a wireless light on a Philips Hue bridge.
First, install the python requirements:
pip install -r requirements.txt
You will then need to download vggish_model.ckpt from AudioSet and move it to the audioset directory to run the vggish model.
You will also need to install Portaudio to run the live inference script. On macOS you can accomplish this with homebrew
$ brew install portaudio
You can run the laugh detector using a pre-trained model from the command line
$ python live_inference.py --save_file='\path\to\output.csv'
If you want to control a Philips Hue bulb, first get the IP address of your bridge, then run
$ python live_inference.py --hue_lights='True' --hue_IP='your.bridge.ip.address'
The first time you run this, you will need to press the connection button within 30 seconds of the phue library loading. See the phue library if you have trouble connecting
The dashboard is built using Dash, you can run it with
$ python dashboard/dashboard.py
It is currently working on a sample data set, but you can edit the file to point to a new datafile.
This repo contains three pre-trained models in the Models directory. You can run the code in Notebooks/ModelTraining.ipynb to repeat the training process or to use it as a skeleton for training your own models. This notebook uses keras-tqdm for the progress bars, which doesn't play well with JupyterLab. You can either run the notebook in a classic Jupyter Notebook or disable keras-tqdm, it won't change the training process.
This repo contains two processed subsets of the AudioSet data: a training set of 18,768 samples where half the data has a laughter label (including "Laughter", "Baby laughter", "Giggle", "Snicker", "Belly laugh", and "Chuckle, chortle") and the other half has a non-laughter human noise label (including "Speech", "Male speech, man speaking", "Female speech, woman speaking", "Child speech, kid speaking", "Conversation", "Narration, monologue", "Babbling", "Crowd", and "Hubbub, speech noise, speech babble"). Samples can have multiple labels, but a positive "laughter" label is only assigned if one of the laughter labels is present and a negative label affirms the absence of any laughter label. The evaluation set contains 568 samples from the evaluation data with the same distribution. These rebalanced datasets are in Data/bal_laugh_speech_subset.tfrecord and Data/eval_laugh_speech_subset.tfrecord respectively.
The pre-trained models and re-balanced subset data are all licensed under a Creative Commons Attribution 4.0 International License.
If you want to create your own processed subset of the AudioSet data on a different collection of labels, you can follow the code in Notebooks/AudiosetProcessing.ipynb. You can create your own subsets of the Audioset data by creating a list of labels to use for the positive and negative cases.