ReplayGain equal loudness filter
Reply #5 – 2010-12-26 20:32:42
Sorry to bump this, but there's been quite some discussion about ReplayGain (RG) lately, so it seems appropriate to me.www.speech.kth.se/prod/publications/files/3319.pdf This is an interesting Master's thesis, a comparative study on file-based loudness estimation algorithms incl. ITU BS.1770 (see David's post) and RG. Here are my main conclusions drawn from the thesis:Exact equal-loudness weighting in the range above ~500 Hz or so seems irrelevant. Some of the best-performing algorithms don't use any weighting above that frequency. RG and BS.1770 both perform very well on average, with BS.1770 producing slightly less errors. IIRC this coincides with the findings reported in this AES 2010 paper . A gating function for ignoring quiet passages in the loudness estimation can improve the performance of BS.1770. See also EBU document 3341 on loudness metering . Since David posted the original RG equal-loudness filter coefficients before 2003, I assume he used the old ISO curves. Now, what to do? I'm also in favor of adopting the BS.1770 algorithm to compute RG values (it's public domain). The remainder of the RG specification could of course stay the same. Moreover, we could try to improve BS.1770 similar to what has been done in EBU R 128 (esp. "integrated mode" in doc 3341 above) and check whether the "Beatles-Fusion/Rock/Metal" loudness mismatch reported by greynol goes away. What do you think? Chris