In an embodiment, a method includes receiving a plurality of first images including at least a mouth-related portion of a human uttering a voice, wherein each first image has depth information; obtaining a noisy spectrogram including a first representation of the voice of the human; extracting a plurality of visual features using the first images, wherein one of the visual features is obtained using the depth information of a second image of the first images; extracting a plurality of audio features using the noisy spectrogram; determining a first spectrogram using the visual features and the audio features; subtracting the first spectrogram from the noisy spectrogram, to obtain a purified representation of the voice of the human; and outputting, by an I/O-related outputting device, a response using the purified representation of the voice of the human.