Thanks to the valuable contributions of George and Daniele, we have made our code for a confomer model with early-exit openly available. This code enables the creation of dynamic neural networks designed specifically for automatic speech recognition.

https://github.com/SpeechTechLab/early-exit-transformer