Back while using a different OS, the author used a program that processed real-time audio to make it sound farther away and lighter. Wanting a simple equivalent for Linux, the author decided to write one. Listen with headphones. The excerpt is from P.I.M.P. by Bacao Rhythm & Steel Band. The processed version sounds farther from the ears and feels less "on the ear," making it feel less "blocked" after prolonged listening.
Since the author was writing for Linux, the base stack was Pipe Wire. With no prior audio experience, an LLM was used as a sounding board while experimenting with standard audio filters and psychoacoustic techniques. The author learned about HRTFs, which describe how sound from a particular direction reaches the ears based on direction and listener geometry. Initial attempts using the MIT KEMAR database sounded awful due to anechoic chamber recording. The SADIE II database, which included room responses, sounded much closer to the remembered effect. However, the approach showed limits as the LLM discussed perceptual envelopment and late-field decorrelation without actionable guidance. The project was left dead for a while.
Later, the author shifted strategy instead of discovering internal parameters. Using a friend's laptop for comparison, the author generated audio on Linux, played it through the program, and recorded the output. Impulse-response measurement was key; a single nonzero sample reveals the filter's decay, shape, and phase delays. The convolution formula (x*h)[n] was applied, assuming the filter is linear time-invariant. A louder impulse produces proportionally louder response, and combined inputs produce separate sums.
Source: Hacker News · Summarized by HeadlinesBriefing