Does WAV sound better than a 320 MP3 on a club system?
Formats · 3 min read
Usually not in a way anyone can hear: in the blind tests we found, trained listeners stop reliably preferring CD audio over MP3 somewhere between 192 and 256 kbps. But no test we found was run in a club, at club level, or on club music. The one difference you can measure on a club rig is not tone. It is how far the decoded MP3 peaks above full scale.
This note runs no test of its own. It reads the published ones, the manufacturers' manuals, and our own 10-track MP3 round trip.
What the listening tests found
Three things repeat:
- The material matters more than the format. One codec, inaudible on some items, audible on others.
- The listener matters. Sound engineers heard more than musicians.
- Electronic music is not automatically easy. One electronic sample split the 2008 test the most.
The club chain treats both files the same
No stage is coarse enough to hide a codec artefact or create one. What changes from the lab is the listener: up to 100 dB(A), a reverberant room, a crowd, maybe earplugs. No test we found covers that, in either direction.
The difference you can measure: peaks
An MP3 stores a description of the waveform, and the decoder rebuilds it. The rebuilt peaks, says the AES streaming recommendation, "may be higher than the peak level at the encoder's input". Club masters leave no room for that: most already sit at or above full scale (Club loudness, measured). On our 10-track round trip, five tracks peaked above 0 dBTP before the MP3 and nine after (What does a 320 kbps MP3 throw away compared to the lossless file?).
Whether that clips depends on the player. Decoders that use fixed-point arithmetic "can clip internally unless designed with sufficient headroom". A clip there would come before the trim knob, so turning the channel down would not undo it. The CDJ-3000 manual does not say what its decoder does. The WAV's own inter-sample peaks meet the converter one step later. And the same AES document notes that limiting by more than about 1 dB "may produce more audible artifacts than simply letting the audio material clip on occasional transients".
So when could the format matter?
- A hot master on a player that clips in its decoder. Audible if the overs are large or frequent.
- Hard material. Nobody has tested 320 on club music.
- Processing after the decode: EQ, key lock, re-encoding a recorded set. Every test played one unprocessed encode.
- A 320 made from something lower.
What this does not tell you
- How it sounds in a club. No test used a club, DJs or club music.
- What we hear. We ran no listening test.
- What a CDJ decoder does with an over. The manual is silent; we did not measure it.
How we checked the sources, with every number
Method. No listening test and no new measurement. Every finding comes from a source below; the peak numbers in the body come from our MP3 round trip (LAME 3.100 at 320 kbps CBR, decoded to 32-bit float, BS.1770 true peak), drawn and tabulated per track in What does a 320 kbps MP3 throw away compared to the lossless file?. content-social/drafts/scripts/I12_figures.py draws the two figures here from the sources below.
Which tests count. A test only counts here if it was blind and compared against the uncompressed original. Two ITU methods matter. ITU-R BS.1116 is built for small impairments, the case of a high-bitrate codec. ITU-R BS.1534 (MUSHRA) is built for intermediate quality, which is why nobody uses it to test a 320.
| Test | Method | Listeners | Material | Playback | MP3 bitrate | Finding |
|---|---|---|---|---|---|---|
| MPEG-2 AAC verification, 1998 | BS.1116, hidden reference | 31 with a professional audio background (22 analysed) | 10 items picked as the most critical of 42 | studio loudspeakers, NHK listening room | 128 (1997 Fraunhofer encoder) | MP3 statistically indistinguishable from the original on 3 of 10 items; AAC at 128 on 8 of 10 |
| Hydrogenaudio public MP3 test, 2008 | ABC/HR, hidden reference, 1-5 impairment scale | volunteers, 26 to 39 per sample | 14 samples known to be hard | headphones, at home | VBR around 128 (sample averages 139 to 144) | All five encoders tied, means 4.26 to 4.59; the low anchor scored 1.56 |
| Pras, Zimmerman, Levitin, Guastavino, AES 2009 | double-blind A/B preference, 150 trials each | 13 trained: 9 sound engineers, 4 musicians | 5 excerpts under 10 s: pop, rock, contemporary, orchestra, opera | loudspeakers in an ITU-standard room | 96 to 320 (LAME) | CD preferred significantly up to 192; no significant preference at 256 or 320 |
| Cunningham, McGregor, 2019 | rating of noise and distortion | 100 untrained (68 valid on that question) | 10 pop excerpts, 20 s | closed headphones at 82 dBC | 192 CBR (and AAC 192) | No significant difference from the uncompressed WAV |
On that 1-5 scale, 5 means "imperceptible" and 4 means "perceptible, but not annoying". So the 2008 averages sit between the two, and that is at 128-class settings on deliberately hard samples. The organiser closed with "This was the last test conducted by me at this bitrate" and proposed moving down to 96 or 80 kbps. We found no comparable public test at 320.
In the tests figure, the Hydrogenaudio test gets its own marker (a square), not "difference found": its result is an average score just under transparent (4.26 to 4.59 of 5), with all five encoders tied against each other, not a significance test against the original. It sits at 141.5 kbps, the middle of its sample averages (139 to 144). The MPEG dot counts as found because MP3 was distinguishable on 7 of 10 items.
The three repeats, in full.
- Material. Both tests that report per-item results found the same codec indistinguishable on some items and audible on others. In the MPEG test, the effect of the item was larger than the effect of the codec.
- Listener. Pras et al. found sound engineers more sensitive than musicians. The Hydrogenaudio page warns that its plots are group averages and that individuals vary.
- Electronic music. In the 2008 test, one electronic sample pushed the VBR encoders up to 192-228 kbps and still produced the widest spread of the 14, with averages from 2.79 to 4.74 depending on the encoder.
The chain, stage by stage.
| Stage | Example | What its manual says |
|---|---|---|
| Player | Pioneer DJ CDJ-3000 | MP3 up to 320 kbps at 44.1/48 kHz, WAV at 16 or 24 bit up to 96 kHz. Analog out 2.0 Vrms; digital out 96 kHz, 24 bit |
| Mixer | Pioneer DJ DJM-900NXS2 | Runs at 96 kHz. 24-bit A/D on the channel inputs, 32-bit D/A on the master |
| Speaker controller | L-Acoustics LA12X | Every input resampled to an internal 96 kHz clock, 32-bit float DSP with IIR and FIR filters, a filter that compensates high-frequency loss in air, and built-in protection against over-excursion, heat and over-voltage |
| Room and level | Swiss V-NISSG | Up to 100 dB(A) averaged over an hour, never above 125 dB(A) |
With analog cables between player, mixer and speaker controller, both files are decoded, converted to analog, converted back at 96 kHz, mixed, converted to analog, then converted again at the speaker controller. None of these stages is coarse: the mixer specifies 105 dB signal-to-noise on its line inputs. The chain does not hide a codec artefact, but it does not create one either. It treats the WAV and the MP3 exactly the same way.
The listener in a club. They are not switching back and forth between short excerpts in a quiet room. They are at up to 100 dB(A), in a reverberant room, among a crowd, perhaps wearing the earplugs Swiss organisers must offer for free above 93 dB(A). The NIOSH recommended exposure limit is 85 dBA over 8 hours with a 3 dB exchange rate. Halve the time for every 3 dB and 100 dBA works out at 15 minutes.
Peaks, in the AES's words. An MP3 does not store the waveform. It stores a description the decoder rebuilds. The AES streaming recommendation says the level at the decoder's output "may be higher than the peak level at the encoder's input". It adds that high-rate coders (its example is 256 kbps) "may work satisfactorily with as little as -0.5 dBTP" of limiter ceiling, and that lower bitrates need more room.
- The AES document warns that "some players have decoders that use fixed-point arithmetic. These decoders can clip internally unless designed with sufficient headroom". A clip inside the decoder happens before the trim knob, so turning the channel down does not undo it.
- The CDJ-3000 manual lists MP3 and AAC at "16 bit" and says nothing about whether its decoder keeps headroom above full scale. From the manual alone, we cannot tell.
- The WAV is not immune. Its own inter-sample peaks meet the player's D/A reconstruction filter, the same issue one step later.
- The same AES document notes that limiting by more than about 1 dB "may produce more audible artifacts than simply letting the audio material clip on occasional transients". Occasional short overs are a measurable fact about the file. They are not proof that the dancefloor hears a worse sound.
Our round trip. Ten club tracks through LAME 3.100 at 320 kbps CBR and back: true peak rose a median +0.35 dB (-0.13 to +1.37); five lossless files measured above 0.0 dBTP, nine decodes. Per-track numbers, the figure and the method are in What does a 320 kbps MP3 throw away compared to the lossless file?.
When the format could matter, at length.
- A hot master plus decoder overshoot, on a player that clips in the decoder. Measurable. Audible only if the overs are large or frequent.
- Hard material. The tests above found audible differences on specific items at 128 kbps, and nobody has published a BS.1116 test of 320 kbps on club music.
- Anything done after the decode: heavy EQ boosts, filter sweeps, key lock, recording the set and encoding it again. Every test in the table played one encode, unprocessed. We found no controlled test of what processing does to coding artefacts.
- A 320 that started life as something lower. The file can only be as good as its source. Where a 320 comes from: What does a 320 kbps MP3 throw away compared to the lossless file?.
More limits, at length.
- No test was run in a club, at club level, with DJs as listeners, or with club music as the main material. Pras et al. used pop, rock, contemporary, orchestral and opera excerpts.
- Encoders change. The MPEG test used a 1997 encoder and the Hydrogenaudio test used LAME 3.97 and 3.98.2. Pras et al. used LAME, version not given in the slides we read. A modern encoder at 320 has, as far as we found, no published BS.1116 result.
- "Not significant" is not "identical". 13 listeners and 5 excerpts can miss a small effect. Pras et al. also measured preference, not whether listeners could tell the files apart.
- Group averages hide individuals. A trained listener with the right sample can sometimes hear what a panel average does not show.
- We did not run a listening test of our own. Everything above is someone else's panel.
- We do not know what a CDJ decoder does with an over. The manual does not say, and we have not measured it.
Sources
- ISO/IEC JTC1/SC29/WG11 N2006, D. Meares, K. Watanabe, E. Scheirer, "Report on the MPEG-2 AAC Stereo Verification Tests", February 1998. https://sound.media.mit.edu/resources/mpeg4/audio/public/w2006.pdf
- S. Mares, "Results of the public MP3 listening test @ 128 kbps", Hydrogenaudio, October 2008. https://listening-tests.hydrogenaudio.org/sebastian/mp3-128-1/results.htm (test description: https://listening-tests.hydrogenaudio.org/sebastian/mp3-128-1/index.htm)
- A. Pras, R. Zimmerman, D. Levitin, C. Guastavino, "Subjective Evaluation of MP3 Compression for Different Musical Genres", AES 127th Convention, paper 7879, 2009. https://aes2.org/publications/elibrary-page/?id=15074 (conference slides: https://www.musicaememoria.com/docs/AES_Presentation_2009.pdf)
- S. Cunningham, I. McGregor, "Subjective Evaluation of Music Compressed with the ACER Codec Compared to AAC, MP3, and Uncompressed PCM", International Journal of Digital Multimedia Broadcasting, 2019, article 8265301. https://doi.org/10.1155/2019/8265301
- Pioneer DJ, CDJ-3000 Operating Instructions (DRI1586-A): supported file formats and specifications. https://www.pioneerdj.com/en/support/documents/
- Pioneer DJ, DJM-900NXS2 Operating Instructions (DRH1330-C): specifications, audio section. https://www.pioneerdj.com/en/support/documents/
- L-Acoustics, LA12X specifications (LA12X_SPS_EN_2.1). https://l-acoustics.com/products/la12x/
- Federal Office of Public Health (BAG), "Schall: Anforderungen der V-NISSG" (Ordinance SR 814.711). https://www.bag.admin.ch/de/schall-anforderungen-v-nissg
- NIOSH, Criteria for a Recommended Standard: Occupational Noise Exposure, Revised Criteria 1998, publication 98-126. https://www.cdc.gov/niosh/docs/98-126/
- Audio Engineering Society, AESTD1008.1.21-9, "Recommendations for Loudness of Internet Audio Streaming and On-Demand Distribution", September 2021, section "Peak Control". https://www.aes.org/technical/documentDownloads.cfm?docID=731
- ITU-R BS.1116-3, Methods for the subjective assessment of small impairments in audio systems. https://www.itu.int/rec/R-REC-BS.1116
- ITU-R BS.1534-3, Method for the subjective assessment of intermediate quality level of audio systems (MUSHRA). https://www.itu.int/rec/R-REC-BS.1534