Transient learning of machine learning models

Embodiments disclosed herein relate to transient learning of a machine learning ("ML") model based on gradient (s) generated at a remote system (e.g., a remote server). A processor (s) of a remote system can receive, from a client device of a user, a stream of audio data (s) that captures spoken utterance (s). The fulfillment pipeline can process the audio data stream (s) such that the particular fulfillment (s) of the spoken utterance (s) is performed. At the same time, the training pipeline can process the audio data stream (s) using an unsupervised learning technique to generate the gradient (s). After processing by fulfilling the pipeline and/or training the pipeline, the audio data stream (s) are discarded by the remote system. Thus, ML model (s) can be trained at a remote system without storing or recording the audio data stream through its non-transitory memory, providing a more efficient training mechanism for training ML model (s) and increasing the security of user data. 本文公开的实施方式涉及基于在远程系统(例如，远程服务器)