SoundCloud upgraded its AAC encoder to Fraunhofer's libfdk_aac, which applies a low-pass filter cutting frequencies above ~17 kHz. While this looks like a downgrade in spectrograms, it's actually an improvement: human hearing is least sensitive above 17 kHz, so dropping those frequencies frees up bits for the mid-range where ears are sharpest. The old encoder spread distortion across audible frequencies and introduced artifacts; the new one makes one deliberate, perceptually cheap sacrifice to improve accuracy everywhere else. The post includes visualizations comparing error rates between encoders and a blind listening test to demonstrate the difference is inaudible.
Table of contents
Lossless vs lossyMake Every Bit CountWhy the Upper Frequencies?What this looks like on real audioBut what about 256 kbps?Listen and DecideThe bottom line23 Impressions