[patrakov]

There is an older project with better math inside: http://drc-fir.sourceforge.net/

For starters, it doesn't try to achieve a phase-neutral response, because a phase-neutral response created in a room is only valid in one point of the room, and creates pre-echo artifacts elsewhere. In fact, it tries to separate the response of the speaker itself from the response of the room, by setting a threshold in the time domain, so that everything coming before it must be unaffected by the room. Then, everything coming before the threshold is corrected to a linear phase, while everything else is corrected to the minimum phase (thus making the second part of the filter purely causal).

Also, they provide an argument, citing literature, that equalizing to a flat frequency response would be wrong in a room, and thus provide an option to remove excessive treble and achieve a 1dB/octave roll-off.

Please see the details at http://drc-fir.sourceforge.net/doc/drc.html

[erdewit]

> because a phase-neutral response created in a room is only valid in one point of the room

Author here. The term "phase-neutral" simply means here that the impulse response is symmetrical and doesn't add a phase shift. It doesn't even try to neutralize the phase characteristics of the room, which is what you may be thinking. In fact the phase information from the measurement is completely discarded. Furthermore, the frequency response is averaged to get a more general and robust (less over-fitted) correction that works pretty well across the room. Try it...

[patrakov]

Well, if you discard the phase of the original response anyway, then you can shave a few milliseconds of latency by switching to minimum-phase (which is causal, not symmetric) instead of linear-phase. The math is in scipy.signal.minimum_phase.