Machine learning models can be trained to cancel noise of diverse types or spectral characteristics, e.g. traffic noise, background chatter, etc. Such models are trained by feeding training data that includes labeled noise waveforms, which is an expensive and time-consuming procedure. Further, the effectiveness of such machine learning models is limited in canceling types of noise absent from training data. Trained models occupy significant amounts of memory which limits their use in consumer devices. This disclosure describes the use of federated learning techniques to train noise canceling models locally at diverse device locations and times. With user permission, the trained models are tagged with timestamp and location, such that when a user device has time or location matching a particular noise cancellation model, the particular model is provided to the user device. Noise cancellation on the user device is then performed with a compact machine learning model that is suited to the time and location of the user device.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Goel, Shantanu and Itankar, Piyush T., "Spatially and Temporally Directed Noise Cancellation Using Federated Learning", Technical Disclosure Commons, (March 19, 2020)