HydrogenAudio

Lossy Audio Compression => Opus => Topic started by: Steve Forte Rio on 2017-10-17 17:09:34

Title: Opus killer @128 kbps
Post by: Steve Forte Rio on 2017-10-17 17:09:34
3 years ago one of the users from my forum published the ABX log (https://audiophilesoft.ru/forum/3-85-1#5017) (russian) which shows that he hear difference between lossless and opus -b 128. As I understand that was opus 1.1 (October 2014)

Now I tried that sample with libopus 1.2.1 (also -b 128) and here's the result:

Spoiler (click to show/hide)

The difference is almost at hearing threshold but is perceptible and sounds like some additional high-frequency noise (Opus sounds with more sharpness than original sample).

The sample: 02_the_riddler_hide_and_seek.flac (https://audiophilesoft.ru/misc/killer-samples/02_the_riddler_hide_and_seek.flac)
Encoded sample: 02_the_riddler_hide_and_seek.opus (https://audiophilesoft.ru/misc/killer-samples/02_the_riddler_hide_and_seek.opus)

As I read in Wiki (https://wiki.xiph.org/Opus_Recommended_Settings#Recommended_Bitrates), Opus recommended bitrate for music is 96—128 kbps, but why it's restricted there to maximum of 128? Maybe it should be raised to ~160 kbps?
Title: Re: Opus killer @128 kbps
Post by: halb27 on 2017-10-17 19:46:38
Would you mind trying --bitrate 140?

(I have 2 tracks I care about which I can't really ABX at --bitrate 128 with 'reasonable effort', but I repeatedly got at 6/8. Too low a safety margin for my taste. Using --bitrate 140 I get real random ABX results.)

The wiki recommendation isn't wise in this respect to me either. We have seen in another thread that --bitrate 80 yields results which are fine usually. Even --bitrate 64 is very usable for a good quality. And if someone wants a setting which is fine also in rare situations -- bitrate 128 isn't a bad choice, but a somewhat higher bitrate maybe useful. What an exact bitrate to use depends on personal preference and personal judgement about those tracks requiring a higher bitrate (relevance and quality wise).
I'd prefer a recommendation for good quality music which reads like 64 ... 160 kbps. (OK, the wiki table is meant to provide a start-off, but I'm not sure whether everybody reads this. Even for a start-off the lowest recommended bitrate should be 80 kbps rather than 96 IMO).
Title: Re: Opus killer @128 kbps
Post by: lithopsian on 2017-10-17 20:40:42
Although you requested 128 kbps, the download segment is actually 175 kbps, presumably trying real hard to be transparent.
Title: Re: Opus killer @128 kbps
Post by: WernerGeidner on 2018-03-03 18:30:14
Sorry to drag up an old thread but I repeated this test with this sample and was clearly able to ABX these two tracks. This system is running on board audio (Realtek® ALC892 codec) and set of cheap Creative T20s. As lithopsian noted Foobar2k reports 175kbps. I repeated the test as per halb27's request at --bitrate 140 and was not able to ABX the tracks. Foobar2k reports a bitrate of 192kbps, however.


foo_abx 2.0.2 report
foobar2000 v1.3.17
2018-03-03 17:34:40

File A: 02_the_riddler_hide_and_seek.opus (bitrate 128)
SHA1: 053283c5c0b788992fa714c5f1a33ca8a559e405
File B: 02_the_riddler_hide_and_seek.flac
SHA1: 1a20bc53d5148ff9963086ba454129a5ed86a676

Output:
WASAPI (event) : Speakers (Realtek High Definition Audio), 16-bit
Crossfading: NO

17:34:40 : Test started.
17:35:23 : 00/01
17:35:42 : 01/02
17:35:53 : 02/03
17:35:59 : 03/04
17:36:23 : 04/05
17:36:37 : 05/06
17:37:32 : 06/07
17:39:13 : 06/08
17:39:39 : 06/09
17:39:53 : 07/10
17:40:17 : 07/11
17:40:55 : 08/12
17:40:55 : Test finished.

 ----------
Total: 8/12
Probability that you were guessing: 19.4%

 -- signature --
56bb08c58d94fba4675383eb6278630dc666cb50


foo_abx 2.0.2 report
foobar2000 v1.3.17
2018-03-03 17:53:56

File A: 02_the_riddler_hide_and_seek.flac
SHA1: 1a20bc53d5148ff9963086ba454129a5ed86a676
File B: 02_the_riddler_hide_and_seek.opus (bitrate 140)
SHA1: aac64935d4575b31d82733021c087d9478ace6ab

Output:
WASAPI (event) : Speakers (Realtek High Definition Audio), 16-bit
Crossfading: NO

17:53:56 : Test started.
17:55:26 : 01/01
17:55:45 : 02/02
17:56:04 : 02/03
17:57:17 : 03/04
17:57:22 : 04/05
17:57:35 : 04/06
17:57:49 : 04/07
17:58:02 : 05/08
17:58:55 : 05/09
17:59:06 : 05/10
17:59:14 : 05/11
17:59:40 : 05/12
17:59:40 : Test finished.

 ----------
Total: 5/12
Probability that you were guessing: 80.6%

 -- signature --
1198c25a82b1c6267a53c45094a34067a0f11d27
Title: Re: Opus killer @128 kbps
Post by: halb27 on 2018-03-04 08:41:43
Thank you for the test.
Because of VBR it's normal that bitrate of an indivual track can deviate a lot from the nominal setting. It just shows in this case that opus knows that the track is hard to encode.

What I'd like to see improved with opus is that average bitrate of many tracks corresponds better with nominal bitrate. Sure average bitrate depends on the test set. My personal test set however corresponds well with average bitrate reported by other people (in various non-opus related situations). But average bitrate of my test set is quite a bit higher than opus nominal bitrate (don't remember the exact percentage - it's not a big deal but also not negligible).
Title: Re: Opus killer @128 kbps
Post by: halb27 on 2018-03-04 08:43:29
Thank you for the test. A good result for moderate opus bitrates.
Because of VBR it's normal that bitrate of an indivual track can deviate a lot from the nominal setting. It just shows in this case that opus knows that the track is hard to encode.

What I'd like to see improved with opus is that average bitrate of many tracks corresponds better with nominal bitrate. Sure average bitrate depends on the test set. My personal test set however corresponds well with average bitrate reported by other people (in various non-opus related situations). But average bitrate of my test set is quite a bit higher than opus nominal bitrate (don't remember the exact percentage - it's not a big deal but also not negligible).
Title: Re: Opus killer @128 kbps
Post by: magicgoose on 2018-04-16 13:06:23
What I'd like to see improved with opus is that average bitrate of many tracks corresponds better with nominal bitrate. Sure average bitrate depends on the test set. My personal test set however corresponds well with average bitrate reported by other people (in various non-opus related situations). But average bitrate of my test set is quite a bit higher than opus nominal bitrate (don't remember the exact percentage - it's not a big deal but also not negligible).
It's a futile goal (everyone's music tastes are different), and also it will shift recommended nominal bitrate as Opus encoder gets more smart in future. (less bits would be needed to reach the same quality)
IMO it would be better to eschew "nominal bitrate" notion for VBR mode and instead use a meaningful normalized quality scale, which would use just as many bits as necessary to reach the quality goal. For example, like this: 
0.0 (default) = do all what it takes to be transparent for approximately 100% of the population (on any good headphones and for all kinds music; killer samples are to be considered as encoder bugs)
1.0 = spend ~2x as many bits than 0.0 would for the same content
-1.0 = spend ~0.5x as many bits than 0.0 would for the same content
etc.
Title: Re: Opus killer @128 kbps
Post by: dhromed on 2018-04-16 13:47:57
clearly able to ABX these two tracks.

Total: 8/12
Probability that you were guessing: 19.4%

I'm not sure if 20% counts as "clearly able". Maybe you can do the test again?
Title: Re: Opus killer @128 kbps
Post by: Simon_ Thunder on 2018-04-16 17:41:15
Its ok with that music style, but try to test jazz and metal songs.