AT 160 KBPS, EVEN THE SHARPEST EARS CAN'T BEAT THE CODEC

Compression with loss is the main way music is distributed and streamed. A codec throws away part of the sound, betting that the human ear will not miss it. Opus, an open-source codec standardised by the Internet Engineering Task Force, is widely used for streaming; it combines one engine designed for speech with another designed for music, and the literature the authors cite finds it better than MP3, Vorbis or HE-AAC at the same bitrate.

Yet most careful listening tests of Opus have focused on speech and low bitrates. Benjamin Martin and Romain Hennequin, of Deezer Research, asked a simple question: from what bitrate does compressed music become impossible to tell from an uncompressed CD-quality original? Community tests had hinted at an answer, they note, but without formal, reproducible results.

A test built to catch small flaws

They used the international standard for detecting small impairments, ITU-R BS.1116-3. In each trial, a listener first hears the uncompressed reference, labelled as such. Then come two blind excerpts, A and B: one is the compressed version, the other a hidden copy of the reference. The listener grades both from 1 (“very annoying”) to 5 (“imperceptible”), and must always give exactly 5.0 to one of them, even if it means guessing. The score that counts is the gap between the compressed version’s grade and the hidden reference’s.

Everything was controlled: isolated booths with ambient noise below 32 decibels, closed-back headphones calibrated flat, sessions of at most 30 minutes with an hour’s rest between them, and a deliberately poor 32 kbps version slipped into every session to check that listeners could hear anything at all.

The music was chosen to be hard: 23 excerpts, including seven worst-case samples from a community benchmark and pieces rich in dense harmonics, such as plucked strings, known to trouble the codec. Twenty-six volunteer employees of Deezer took part, the authors excluded; five failed the screening — two for breaking the rules, three because they could not reliably detect the 32 kbps version — leaving 21.

128 passes, then gets caught

In the first phase, 32 and 64 kilobits per second were clearly detected. 128 kbps was not: its average gap of −0.08 points was statistically indistinguishable from zero (p = 0.456). It narrowly failed, however, a stricter test that asks for positive proof of equivalence within ±0.25 points.

So the team built a worst case. Using a statistical model that separates how critical each listener is from how revealing each excerpt is, they kept only the 10 most critical listeners and the 8 excerpts whose artefacts were easiest to hear, and tested again at 32, 96, 128 and 160 kbps. Under that scrutiny, 128 kbps was detected (gap −0.24, p = 0.030).

160, indistinguishable

At 160 kbps, the gap fell to −0.04 (p = 0.747): these expert ears no longer told it apart from the original. And this time the stricter test passed too (p = 0.024), giving positive statistical evidence of equivalence.

The authors’ conclusion is two-tiered. For a general audience, 128 kbps already shows no significant degradation. For critical listening on the hardest material, 160 kbps is needed to reach perceptual transparency. They are careful about its scope: one model of headphones, a panel drawn from a single company, and only ten listeners in the decisive phase. Testing with other audiophile equipment and a broader pool of experts, they write, would be needed before calling the threshold universal. Their results, anonymised listener profiles and excerpt characteristics are publicly released for anyone who wants to try.