Abstract: Bandwidth-efficient (i.e., compressed) representations of audio data can be utilized for near real-time presentation of the audio on one or more receiving devices. Persons identified as having speech represented in the audio data can have trained speech models provided to the devices. These trained models can be used to classify the compressed audio in order to improve the quality to correspond more closely to the uncompressed version, without experiencing lag that might otherwise be associated with transmission of the uncompressed audio. The uncompressed audio is also received, with potential lag, and is used to further train the speech models in near real time. The ability to utilize the uncompressed audio as it is received prevents a need to store or further transmit the audio data for offline processing, and enables the further trained model to be used during the communication session.

Abstract: Techniques for using machine learning models to approximate a user in a communication are described. For example, a method of initiating a communication link with an edge device to exchange audio data; receiving a compressed audio data from the edge device; re-encoding the received compressed audio data using a re-encoding machine learning model to approximate a voice; and outputting the approximated voice is detailed.