Lossless vs lossy audio, and the re-encode nobody counts

The distance between a competently made lossy file and the lossless master it came from is smaller than the distance your playback chain adds afterwards, and on wireless headphones the lossless file is decoded and re-encoded into a lossy one before it reaches the driver. The published listening tests support a narrower claim than either side of this argument usually makes, and the ITU recommendations those tests run under spell out exactly why.

What a lossy encoder throws away

Recommendation ITU-R BS.1387-2 (May 2023), the ITU's method for objective measurement of perceived audio quality, states the design principle in one sentence. Data reduction algorithms "are adapted to the properties of the human auditory system and particularly rely on masking effects. Such algorithms do not aim mainly at minimizing the distortions but rather attempt to handle these distortions in a way that they are perceived as little as possible."

The magnitude of that effect is the whole argument. The same document cites what it calls "the so-called 13 dB miracle": "Superimposed noise with a spectral structure adapted to that of the audio signal is almost inaudible even if the resulting unweighted S/N declines to 13 dB." A 13 dB signal to noise ratio is catastrophic by any conventional measurement. Shaped to sit under the masking curve, it approaches inaudibility.

Two things follow. A lossy decoder does not hand back the input waveform, so waveform differences say almost nothing about audibility. And, as BS.1387-2 puts it, "the evaluations of perceptual codecs require listening tests in order to assess the audio quality". No measurement substitutes for a panel of ears.

Where the current encoders sit

RFC 6716 (September 2012) defines Opus across "all bitrates from 6 kbit/s to 510 kbit/s", and its own bitrate guidance for 20 ms frames lists "64-128 kbit/s for FB stereo music", FB meaning fullband. That is the band where the community tests were run, and the reason is instructive: differences there are large enough to resolve. The Opus project's comparison page summarises ABC/HR tests on 48 kHz stereo music at 96 kb/s showing "Opus outperforming two LC-AAC encoders, libvorbis, and a 136 kb/s MP3 encoder". The IETF's own summary of the codec's evaluation record, draft-ietf-codec-results-00, lists MUSHRA panels of 9 to 21 listeners, all at bitrates from 11 to 128 kbps.

Above roughly 128 kbit/s for stereo music the published record thins out, and that is not an accident of interest.

What blind tests are allowed to establish

Two ITU recommendations divide the work, and they are explicit about not overlapping.

ITU-R BS.1534-3 (October 2015) defines MUSHRA, the "Method for the subjective assessment of intermediate quality level of audio systems". Its considerations state "that Recommendation ITU-R BS.1116 is intended for the assessment of small impairments and is not suitable for assessing systems with intermediate audio quality". The reverse also holds. BS.1534-3 warns that "If MUSHRA is used with appropriate content, it is ideal that listener scores should range between 20-80 MUSHRA points. If scores for the majority of test conditions fall in the range of 80-100 it may be true that the results of the test are invalid." Codecs that are nearly transparent pile up at the top of the MUSHRA scale, which invalidates the instrument rather than proving the codecs identical.

ITU-R BS.1116-3 (February 2015) is the method for small impairments, and its demands are steep. Data "should come exclusively from subjects who have expertise in detecting these small impairments". The prescribed method is "double-blind triple-stimulus with hidden reference", graded on a continuous five-grade scale where 5.0 is "Imperceptible" and 4.0 is "Perceptible, but not annoying". Because "long- and medium-term aural memory is unreliable", the procedure "should rely exclusively on short-term memory" with near-instantaneous switching, and "A grading session should not last for more than 20-30 min".

The transparency rule matters most. BS.1116-3 does not let you declare a codec transparent because nobody heard a difference. It requires anchors "known, (e.g. from previous research), to be detectable to expert listeners but not to inexpert listeners" to be planted in the same test. Only if those anchors are correctly identified is apparent transparency "evidence for 'true transparency'". Otherwise "the apparent transparency of systems cannot be properly interpreted, and the experiment will need to be run again". A null result from an untrained panel in an uncontrolled room establishes nothing.

What lossless guarantees, and what it costs

RFC 9639 (December 2024) standardised FLAC after two decades of de facto use. Its guarantee is total and boring: it compresses "losslessly, i.e., it does so without losing information", so that "decompressing losslessly compressed information returns exactly the original data". The reference implementation is FLAC 1.5.0, released 11 February 2025. Every stream carries an MD5 signature of the unencoded audio in its STREAMINFO block, so the claim is checkable rather than asserted. Apple's ALAC gives the same guarantee inside Apple Music, which offers Lossless at 24-bit/48 kHz and Hi-Res Lossless at 24-bit/192 kHz.

The size cost is variable by design. The FLAC FAQ puts the honest range plainly: "the result can be from around 100% of the input rate (if you are encoding noise), down to almost 0 (encoding silence)". A published two-track measurement using FLAC 1.3.0 landed at 69.8% of the source WAV for a dense progressive rock track and 49.0% for a sparse choral recording at compression level 8. CD stereo PCM runs at 1,411 kbit/s (16 bits by 44,100 samples by 2 channels), so those ratios put typical FLAC somewhere near 700 to 1,000 kbit/s against the 320 kbit/s of Spotify's Very High tier. Call it two to three times the storage and the data.

That is a real cost for a benefit that is real but narrow: an archival master you can re-encode any number of times without generation loss, which is exactly what a lossy file cannot give you.

24-bit and 192 kHz for playback

Higher numbers on the delivery format do not do what the store page implies. Xiph.Org's writeup on the subject is blunt about it: "there is no point to distributing music in 24-bit/192kHz format. Its playback fidelity is slightly inferior to 16/44.1 or 16/48, and it takes up 6 times the space." The two mechanisms are specific. With shaped dither, "the effective dynamic range of 16 bit audio reaches 120dB in practice, more than fifteen times deeper than the 96dB claim", so 24-bit adds headroom that playback never uses. And reproducing ultrasonic content through the same transducer as the audible band means "any nonlinearity will shift some of the ultrasonic content down into the audible range as an uncontrolled spray of intermodulation distortion products".

The listening evidence points the same way without being unanimous. Meyer and Moran, in the Journal of the Audio Engineering Society volume 55 issue 9 (September 2007), inserted a 16-bit/44.1 kHz analogue to digital to analogue loop into high resolution playback and reported that "the CD-quality A/D/A loop was undetectable at normal-to-loud listening levels, by any of the subjects, on any of the playback systems". Across 554 trials, listeners "chose correctly 49.8% of the time".

The counterweight deserves stating fairly. Joshua Reiss, in JAES volume 64 issue 6 (June 2016), pooled eighteen published experiments with over 400 participants and more than 12,500 trials and found "a small but statistically significant ability of test subjects to discriminate high resolution content", with the effect growing substantially after training. So the difference is not exactly zero. It is small, it requires training to surface, and it is the wrong order of magnitude to matter against anything else in this article. Note also that 24-bit is genuinely useful during recording and mixing, where headroom before clipping is a working constraint. That is a production argument, not a distribution one.

The re-encode between the file and your ear

Here is where the whole debate usually goes wrong. Bluetooth audio distribution runs over A2DP, currently at profile version 1.4, and every A2DP link carries a lossy codec. The IETF payload draft for SBC states the baseline: "To ensure interoperability, the SBC codec has been specified, in appendix B of the A2DP specification, which shall be included into all A2DP Bluetooth devices", with "Recommended operational modes range from 127 to 345 kb/s". Apple says the same thing about its own hardware in one line: "Bluetooth connections aren't lossless."

So a FLAC or ALAC file on a phone is decoded to PCM and then re-encoded by whatever codec the link negotiated. The lossless guarantee ends at that boundary, every time.

The LDAC ladder is not marketing spin, it is in the Android Bluetooth stack as constants: A2DP_LDAC_QUALITY_HIGH is annotated "Equal to LDACBT_EQMID_HQ 990kbps", and A2DP_LDAC_QUALITY_ABR carries the comment "ABR mode, range: 990,660,492,396,330(kbps)". Adaptive is the default, so the rate you actually get is set by radio conditions in the room, not by the file.

The scale of the impairment is documented by the Bluetooth SIG itself. Testing LC3 against SBC under ITU-R BS.1116-3, it reports that "If you look at the 345-kbps bit rate for example, you'll see that SBC scores a value just above 4.0, whereas LC3 scores an even higher amount with less than half the bit rate (160 kbps)". On the BS.1116 scale, just above 4.0 is "Perceptible, but not annoying". The mandatory Bluetooth codec, at its top recommended rate, sits at an impairment level a trained panel can hear. That is an entire grade point below the ceiling a lossless argument is fighting over.

The room you are listening in

BS.1116-3 also specifies the conditions under which a trained listener is expected to resolve a small impairment at all. Continuous background noise in the listening area "should preferably not exceed NR 10", and "Under no circumstances should the background noise exceed NR 15". Early reflections arriving within 15 ms of the direct sound must be attenuated "in the range 1-8 kHz by at least 10 dB". Floor area for two-channel stereo is 20 to 60 square metres, with reverberation time set by Tm = 0.25 (V / V0)^(1/3) seconds.

That is the environment the codec differences in this article were measured in. A kitchen, a train carriage, or a living room with a soundbar under the television is not that environment, and the gap between them is larger than every codec decision above it.