Convolutional neural network with positional embeddings for audio processing

A method, comprising:

• receiving audio data via an interface for an audio processing system;
• concatenating, by the audio processing system, different vectors of respective frequency positional embeddings centered at different respective time-frequency bins to the different respective time-frequency bins when processing the audio data at an input layer of a convolutional neural network;
• and
• performing, by the audio processing system, an audio processing task based, at least in part, on a result of processing the audio data through the convolutional neural network.

One or more non-transitory, computer-readable storage media, storing program instructions that when executed on or across one or more computing devices cause the one or more computing devices to implement an audio processing pipeline, wherein the audio processing pipeline implements:

• receiving audio data;
• performing a transform on the audio data to determine a plurality of different spectrum bands of the audio data;
• processing the plurality of different spectrum bands of the audio data through a convolutional neural network that concatenates different respective vectors of frequency positional embeddings centered at respective time-frequency bins at an input layer to the different respective time-frequency bins for the plurality of different spectrum bands;
• and
• performing an audio processing task based, at least in part on, a result of processing the audio data through the convolutional neural network.