Artificial intelligence (AI) based garbled speech elimination

A method for processing at least one speech frame, the method being performed by a network node and comprising:

• receiving a first ciphered speech frame comprising first voice-data, wherein the first voice-data comprises encoded voice data;
• deciphering the first ciphered speech frame using a first deciphering parameter, thereby obtaining a first deciphered speech frame comprising the first voice-data, wherein the first deciphered speech frame is an encoded speech frame and comprises a set of information elements (IEs), wherein each IE consists of a set of one or more bits;
• inputting a subset of the set of IEs of the first deciphered speech frame into a trained machine learning (ML) model, wherein the ML model was trained offline using supervised learning, and the ML model was trained using a first version of an audio file comprising GS and a second version of the audio file not comprising GS;
• obtaining from the ML model a first garbled speech (GS) indicator, wherein the first GS indicator indicates either i) that the ML model has detected GS based on the subset of IEs of the first deciphered speech frame or ii) that the ML model has not detected GS based on the subset of IEs of the first deciphered speech frame;
• determining whether the first GS indicator indicates that the ML model has detected GS based on the subset of IEs of the first deciphered speech frame;
• after determining that the first GS indicator indicates that the ML model has detected GS based on the subset of IEs of the first deciphered speech frame, deciphering the first ciphered speech frame using a second deciphering parameter instead of the first deciphering parameter, thereby obtaining a second deciphered speech frame comprising a set of IEs;
• based at least in part on a set of two or more GS indicators obtained from the ML model, determining whether or not a GS condition exists;
• and
• as a result of determining that the GS condition exists, deciphering a subsequent ciphered speech frame using an adjusted deciphering parameter.

A network node for processing at least one speech frame, the network node comprising:

• a receiver for receiving a first ciphered speech frame comprising first voice-data, wherein the first voice-data comprises encoded voice data;
• and
• processing circuitry programmed to cause the network node to perform a process that comprises:
• deciphering the first ciphered speech frame using a first deciphering parameter, thereby obtaining a first deciphered speech frame comprising the first voice-data, wherein the first deciphered speech frame is an encoded speech frame and comprises a set of information elements (IEs), wherein each IE consists of a set of one or more bits;
• inputting a subset of the set of IEs of the first deciphered speech frame into a trained machine learning (ML) model, wherein the ML model was trained offline using supervised learning, and the ML model was trained using a first version of an audio file comprising GS and a second version of the audio file not comprising GS;
• obtaining from the ML model a first garbled speech (GS) indicator, wherein the first GS indicator indicates either i) that the ML model has detected GS based on the subset of IEs of the first deciphered speech frame or ii) that the ML model has not detected GS based on the subset of IEs of the first deciphered speech frame;
• determining whether the first GS indicator indicates that the ML model has detected GS based on the subset of IEs of the first deciphered speech frame;
• and
• after determining that the first GS indicator indicates that the ML model has detected GS based on the subset of IEs of the first deciphered speech frame, deciphering the first ciphered speech frame using a second deciphering parameter instead of the first deciphering parameter, thereby obtaining a second deciphered speech frame comprising a set of IEs.