1.01j encoder vs 1.15u
Reply #1 – 2005-06-18 06:42:13
• samples : 15 classical samples, used on several previous listening tests (multiformat, AAC, Vorbis and LAME), and coming from a selection of best technical & artistic recordings of the years made by a French magazine. • hardware : — Creative Audigy2 soundcard with official driver — Beyerdynamic DT-531 soundcard — Onkyo amp • ABX conditions : — schnofler amazing ABC/HR software (0.5 alpha 5) — fixed number of trials: eight; sixteen if problems (results are hidden until the end of the test) — no blind ABX sessions between reference file and encoded one ; ABX comparison between 1.01j and 1.15u encodings only ; REFERENCE is nevertheless accessible for playback during ABX session in schnofler’s software (so, please, don’t complaint on the theme of “you’ve rated the most pleasant to your ears and not the one which was closest to reference ”) — first second was systematically discarded from evaluation and ABX comparison (if not, it was a mistake; see log files to check it: the listening range is stored) — consequent rests between most tested samples (4h30 to test the complete set of 15 samples) — a lot of coffee, water, soup and even ice cream • Encoders and Decoder : — mppenc.exe 1.01j from (file date) 2002.04.02 (YYYY.MM.DD) — mppenc.exe 1.15u from 2005.02.23 — mppdec.exe 1.93h from 2002.10.22 (default settings of ABC/HR 0.5a5)[span style='font-size:14pt;line-height:100%']RESULTS [/span] ABX logs are here . Surprisingly, there’s an audible regression with latest encoder when compared to an older one. Statistically, it’s apparently safe to say that – to my ears and on these samples which are not really killers – mppenc 1.01j is better than 1.15u. Bitrate can’t be invoked to explain this difference: 1.01j encodings are systematically smaller than 1.15u (12 kbps, corresponding to a +6.4% inflation for latest encoder). So what happened? From what kind of problems suffers 1.15u? First, it should be important to keep in mind that I haven’t rated any sample 5.0 ; none was transparent, and all of them revealed at best minor issues with both encoders. 1.15u doesn’t introduce new artifact, but rather increase the level of audible distortions. But what form of distortion? It’s hard for me to give a precise description of what I’ve heard. But one issue appeared several times. • rinaldo.wav : “now, it's distorted, fluctuant, a bit acid too ” • hercules.wav : “aggressive ringing on some parts ” • bach BWV 1014.wav : “flute is not clean, and suffers from asperities (irregular or unconstant line), like a very subtle ringing ” • compostelle.wav : “It suffers from something like chirping. Reference sound is very tonal, continuous, regular and smooth; encoded one is very messy, irregular, inconstant, like millions bees flying in the church. Weird, and aggressive ” • Bruhns.wav : “Irregular organ note (now…); aggressive ringing. ” Ringing, fluctuant, irregular, inconstant… On five samples, I’ve described the problem as being a form of ringing Other problem: • Brahms : “ slightly lifeless (missing noise). Pre-echo is perceptible ” • Van Wilder : “ pre-echo appears clearly ; sound is also a bit synthetic and lifeless, probably a consequence of missing noise or excessive lowpassing process. But pre-echo is clearly the biggest issue on this encoding. ” Here, something like an excessive lowpass (probably) or ATH threshold lead to a synthetic sound (very subtle) and also pre-echo issue. I’ve encountered similar problems when I’ve tested some problematic versions of lame: aggressive noise reduction was often correlated to higher level of pre-echo (and of course it involves the feeling of synthetic, lifeless, synthetic, claustrophobic ambiance). For other samples, I couldn’t be more precise (acidity, false or wrong color, and of course distortions are filling the log files). On the other side, 1.01j appeared to sound worse on two samples: • Mozart – Dies Irae : sibilant consonant (“s”) were distorted (clear difference) • Avison : the solo violin appeared as more “acid” (subtle difference)LINKS Samples are as usual uploaded for complementary tests (highly welcome). • not available anymore for download mppenc.exe (v.1.01j) could be downloaded here