Audio processing apparatus and method for denoising a multi-channel audio signal

An audio processing method, comprising:

• receiving a respective plurality of audio frames of an audio signal from an audio source, wherein for each audio sensor of the plurality of audio sensors, the respective plurality of audio frames for the audio sensor defines an audio channel of the audio signal;
• for each audio channel of the audio signal, determining a respective feature set having at least one feature for each audio frame of the respective plurality of audio frames for the audio channel, wherein the respective feature sets corresponding to the audio channels define a three-dimensional feature array;
• processing the three-dimensional feature array using a neural network, wherein the neural network comprises a self-attention layer configured to process a plurality of two-dimensional sub-arrays of the three-dimensional feature array to generate a plurality of processed two-dimensional sub-arrays;
• extracting phase information from one audio channel of a plurality of audio channels of the audio signal;
• and
• generating a denoised output audio signal based on the plurality of processed two-dimensional sub-arrays and the phase information.

A non-transitory computer-readable storage medium storing program code that, when executed by a processor, causes a computer to perform steps comprising:

• receiving a respective plurality of audio frames of an audio signal from an audio source, wherein for each audio sensor of the plurality of audio sensors, the respective plurality of audio frames for the audio sensor defines an audio channel of the audio signal;
• for each audio channel of the audio signal, determining a respective feature set having at least one feature for each audio frame of the respective plurality of audio frames for the audio channel, wherein the respective feature sets corresponding to the audio channels define a three-dimensional feature array;
• processing the three-dimensional feature array using a neural network, wherein the neural network comprises a self-attention layer configured to process a plurality of two-dimensional sub-arrays of the three-dimensional feature array to generate a plurality of processed two-dimensional sub-arrays;
• extracting phase information from one audio channel of a plurality of audio channels of the audio signal;
• and
• generating a denoised output audio signal based on the plurality of processed two-dimensional sub-arrays and the phase information.