US 12,170,096 B2
Upsampling of audio using generative adversarial networks
Rithesh Kumar, Montreal (CA); and Kundan Kumar, Montreal (CA)
Assigned to Descript, Inc., San Francisco, CA (US)
Filed by Descript, Inc., San Francisco, CA (US)
Filed on Sep. 17, 2021, as Appl. No. 17/478,722.
Claims priority of provisional application 63/083,474, filed on Sep. 25, 2020.
Prior Publication US 2022/0101872 A1, Mar. 31, 2022
Int. Cl. G10L 25/18 (2013.01); G06F 3/16 (2006.01); G06N 3/045 (2023.01); G06N 3/088 (2023.01); G10L 19/02 (2013.01); G10L 25/30 (2013.01)
CPC G10L 25/18 (2013.01) [G06F 3/165 (2013.01); G06N 3/045 (2023.01); G06N 3/088 (2013.01); G10L 19/02 (2013.01); G10L 25/30 (2013.01)] 15 Claims
OG exemplary drawing
 
1. A method for implementing a generative adversarial network, the method comprising:
receiving, by a processor, input that is indicative of a selection of a first discrete audio signal to be upsampled from a first sampling rate to a second sampling rate;
examining, by the processor, a database that includes multiple generative adversarial networks associated with different sampling rates, so as to identify the generative adversarial network that is associated with the second sampling rate;
applying, by the processor, a transform to the first discrete audio signal to produce a first magnitude spectrogram;
providing, by the processor, the first magnitude spectrogram to the generative adversarial network as input so as to produce a second magnitude spectrogram,
wherein the second magnitude spectrogram is generated from the first magnitude spectrogram by the generative adversarial network by adjusting a characteristic learned, during training, from analysis of multiple magnitude spectrograms, each of which is associated with a different discrete audio signal having the second sampling rate; and
applying, by the processor, an inverse transform to the second magnitude spectrogram to produce a second discrete audio signal having the second sampling rate.