System and method for background noise suppression by projecting an input audio into a higher dimension space

A method for background noise suppression, the method implemented by a background noise suppression system and comprising:

• projecting frames fragmented from input audio data into a higher dimension space than a lower dimension space of the input audio data, wherein the input audio data comprises speech content;
• applying a speech mask to the frames, fragmented from the input audio data and projected into the higher dimension space, to separate speech components and noise components of the frames, wherein the speech components are associated with the speech content;
• and
• transforming the speech components of the frames in the higher dimension space into a feature domain of the input audio data having the lower dimension space by performing a first inverse projection on the speech components to generate output audio data comprising a noise-suppressed version of the input audio data;
• transforming the noise components of the frames in the higher dimension space into the feature domain of the input audio data having the lower dimension space by performing a second inverse projection on the noise components to generate noise;
• and
• training a model, using the generated noise, to distinguish between speech and other noise.

A non-transitory computer-readable medium comprising instructions for background noise suppression that, when executed by at least one processor, cause the at least one processor to:

• project frames fragmented from input audio data obtained via an audio interface into a higher dimension space than a lower dimension space of the input audio data;
• apply a speech mask to the frames, fragmented from the input audio data and projected into the higher dimension space, to separate speech components and noise components of the frames in the higher dimension space;
• transform the speech components of the frames in the higher dimension space into a feature domain of the input audio data having the lower dimension space by performing a first inverse projection on the speech components to generate a noise-suppressed version of the input audio data;
• and
• transform the noise components of the frames in the higher dimension space into the feature domain of the input audio data having the lower dimension space by performing a second inverse projection on the noise components to facilitate training a model to learn a difference between other speech and noise.