Basis is a proof-of-concept web app which provides an interactive demonstration of separating on/off-screen audio sources for a given video.
It leverages the speech separation model created by Andrew Owens et al. used for separating on / off-screen audio sources. This project is based upon open-source code and models licensed under the Apache License 2.0.
I built this as an opportunity to learn:
- Implementation details of a "legacy" TensorFlow 1.x model
- How to freeze, inspect, and host a model using TF-Serving
- How to perform real-time inferencing on video from a public web app