Skip to main content
Topic: Opus 1.3-rc (Read 15277 times) previous topic - next topic
0 Members and 2 Guests are viewing this topic.

Re: Opus 1.3-rc

Reply #50
as far as speech goes... 14kbps is about the lowest I would suggest as while you can go lower and have it still passable (i.e. still understand what's said etc) I feel sound quality starts to really drop off going below the 14kbps setting, at least based on my Klipsch Pro-Media speakers for the PC. because while I can notice the 14kbps being a hit to the overall sound compared to the 128kbps original MP3 speech file I got, it's not that much worse for a tiny fraction the original files bit rate.
Well, it is possible that I may be exaggerating things, but I was in no way saying that I'd listen to a 9kbps Opus song for daily purposes--I just meant that for 9kbps, the result is (at least for me) exceptional compared to what other codecs can offer at the same bitrate. But no, I'd never listen to a 9kbps Opus song for daily listening stuff--for that, I'd happily use (at least) 48kbps (but I'd rather use 64kbps) (I'm not that picky with quality, and a bit of artifact adds a nice touch to the audio, I dunno why I think that way) But 9kbps... never XD

Re: Opus 1.3-rc

Reply #51
Well, it is possible that I may be exaggerating things, but I was in no way saying that I'd listen to a 9kbps Opus song for daily purposes--I just meant that for 9kbps, the result is (at least for me) exceptional compared to what other codecs can offer at the same bitrate. But no, I'd never listen to a 9kbps Opus song for daily listening stuff--for that, I'd happily use (at least) 48kbps (but I'd rather use 64kbps) (I'm not that picky with quality, and a bit of artifact adds a nice touch to the audio, I dunno why I think that way) But 9kbps... never XD

I was just strictly saying 14kbps is THE minimum I would suggest for 'speech' (not music) as if you got lengthy speech files it's a nice way to keep the storage space very minimal for hours and hours of speech while maintaining a decent level of speech quality. I suspect that minimum suggested kbps setting will vary a bit from person-to-person but just testing on my Klipsch Pro-Media speakers that's the conclusion I came to as the sound clarity starts to take a solid hit below 14kbps which is why I drew the line at 14kbps as my minimum suggested setting for speech.

but yeah, considering speech quality of Opus at 9kbps etc, while there is clearly a hit to the sound quality, it's respectable given the VERY low/next to nothing bit rate. but as you said, I am more concerned with sound quality at that point to where I would rather lose a bit of disk space to clean up the sound a bit. hence, my 14kbps+ suggestion.

but for music it's a whole different ball game to where I am pretty much inline with you with the 48-64kbps thing being about the minimum I would possibly use. like for bit rates lower than 96kbps (I use 96kbps as the minimum for someone who's concerned with quality), 64kbps is a solid choice. for those who prefer to use minimal storage space for music, 48kbps is a possible option. hell, even 32kbps is solid considering the very low bit rate but unless someone is trying to cram a bunch of music into the least amount of storage space possible I would generally stick to at least 48kbps+ minimum. but for the most part I stick with 96kbps on more important music for on-the-go listening and for less important music I use 64kbps (using this combo I managed to fit pretty much my entire music collection onto a 8GB MicroSD card). it's possible I would consider 48kbps but it would have to more-or-less be one of those situations where your really hard up for space and I don't want to cut out songs. but come to think of it... I really need to do a little testing to see how music with Opus @ 32kbps/48kbps sounds in comparison to 64kbps on my headphones to see if I can detect any more obvious differences or not as if not, then 48kbps could be a realistic choice for me for further efficiency if I need to squeeze more songs on that 8GB MicroSD card in the future.
For music (especially on-the-go)...
-I suggest Opus @ 96kbps (or... 64kbps minimum, 128kbps maximum). *preferred choice*
-I suggest AAC(Apple) @ 96kbps (q45 TVBR) or 128kbps (q64 TVBR). *secondary choice*
-I use Foobar2000 (/w Encoders Pack etc) to convert FLAC to Opus/AAC(Apple).

Re: Opus 1.3-rc

Reply #52
I was just strictly saying 14kbps is THE minimum I would suggest for 'speech' (not music) ...
Now You can go even lower with 1.3 on speech. As low as 13.2 kbps with even better quality (see below).



  • Using wideband encoding down to 9 kb/s

As far as I can tell, this is due to three different modes being used: SILK narrowband up to 9 kb/s, SILK wideband (the slowest) between 9 and 13 kb/s, then CELT from there on. ...
Now 1.3RC1 encodes to SWB/FB at 13.2 kbps as well ( while 1.2.1 codes to WB at this bitrate).
I've noticed it uses SWB or FB on a different samples. I guess it's expected and desired behavior (?)

Overal quality is better actually for 1.3RC.  Though German speech sample was hard to encode. I wonder if it's matter of training and/or tuning.

Samples: https://drive.google.com/file/d/1BMrQlEz01gJVP43PAtYBYiLQUMHlhsLq/view?usp=sharing



Conclusion:
1.3RC1@13.2kbps is on par with 1.2.1@16kbps on speech.
That's 20% of bitrate reduction.  Similar bitrate reduction is observed on higher bitrates (16-40+ kbps)



Re: Opus 1.3-rc

Reply #53
I find the aggressive push for higher bandwidths at lower bitrates slightly surprising, especially for speech.

I mean, obviously one wants to quit doing narrowband as soon as possible; it misses almost all the energy in many consonants, harming intelligibility, and everything including vowels sounds flat. Mediumband is much better but still an obvious degradation. But wideband is quite good and I'm happier with an 8KHz lowpass than I am with most kinds of distortions. SWB speech is frequently non-ABXable from fullband, since above 12K there's so often little other than noise of dubious audibility.

So it's surprising to find that the encoder will now be spending bits at rather low bitrates on coding 12-20KHz rather than spending more bits on accuracy in the frequency range where the formants and most of the consonant energy are. I wonder what kind of listening tests etc suggested making such a move.

Down the road I wonder whether something like the WaveNet codec2 decoding demo will end up making SBR, hybrid mode, etc mostly obsolete for speech.

Re: Opus 1.3-rc

Reply #54
12-20 kHz frequency range is important for noisy speech and music (stereo material, etc..) http://oh3tr.fi/~oh3gdd/Publications/%5B12%5D_Dallas_ICASSP2010_Ramo_Voice%20Quality%20Evaluation%20of%20Various%20Codecs.pdf

Also Opus codes 12-20 kHz range  in just 2 freq. bands of total 20  (10% of total bands). Add to that band folding and 12-20 kHz range is coded just at few-to-several hundreds of  bits ( less than 1 kbps). 

So it's very cheap for Opus to code this range starting from somewhere ~13.2 -14 kbps.

P.S. Also I suspect that SWB can be pushed down to ~11-12 kbps (at least for speech) because 1.3RC SWB@13.2kbps is on par with an older 1.2.1 FB@16kbps   :o


Re: Opus 1.3-rc

Reply #56
I tried this - opus-tools-test-1.3-rc-fix.zip -
It's almost twice as efficient as aac Itunes . Obviously it has different types of artifacts compared to aac Itunes.
It's not bad I think. When the full 1.3 Opus will come out? Probably in July this year?


Re: Opus 1.3-rc

Reply #58
How does the audio quality of Opus 1.3 at 32 kbps stereo compare to 48 kbps FDK these days?  Testers of earlier versions claimed they were equivalent, but I found that not yet to be the case,

Re: Opus 1.3-rc

Reply #59
Conclusion:
1.3RC1@13.2kbps is on par with 1.2.1@16kbps on speech.
That's 20% of bitrate reduction.  Similar bitrate reduction is observed on higher bitrates (16-40+ kbps)

I was briefly playing around with a speech file I got online (my original copy is a 87kbps VBR MP3(tool in Foobar2000 says 'Lavc56.28')) with v1.3RC vs v1.2.1 I noticed I could lower from 14kbps to 13kbps with v1.3RC and things seemed pretty much the same but going to 12kbps and there was a noticeable hit. although bit rates did not exactly scale down the same from 14kbps to 13kbps to 12kbps...

selected bit rate = actual bit rate...
12kbps = 12kbps (noticeable hit to sound quality vs the two below)
13kbps = 14kbps (seemed to be similar to the setting below)
14kbps = 15kbps (seemed to be similar to the setting above)

basically v1.2.1 @ 14kbps (15kbps real bit rate) is better than v1.3RC @ 12kbps (12kbps real bit rate) on the speech file I tested as the main thing is the v1.3RC @ 12kbps is more muffled where as the v1.2.1 @ 14kbps is not (like the v1.2.1 file @ 14kbps sounds closer to the 87kbps MP3 source file than the v1.3RC @ 12kbps does). is this to be expected?

p.s. this was all done on my Klipsch Pro-Media speakers on the PC with Foobar2000 v.1.3.19.
For music (especially on-the-go)...
-I suggest Opus @ 96kbps (or... 64kbps minimum, 128kbps maximum). *preferred choice*
-I suggest AAC(Apple) @ 96kbps (q45 TVBR) or 128kbps (q64 TVBR). *secondary choice*
-I use Foobar2000 (/w Encoders Pack etc) to convert FLAC to Opus/AAC(Apple).

Re: Opus 1.3-rc

Reply #60
It's expected. Everything lower than 13 kbps is coded to WB(8kHz). Higher than 13 kbps is coded to SWB/FB (12 kHz/20kHz).

That's why earlier I've posted
P.S. Also I suspect that SWB can be pushed down to ~11-12 kbps (at least for speech) because 1.3RC SWB@13.2kbps is on par with an older 1.2.1 FB@16kbps   :o

Re: Opus 1.3-rc

Reply #61
OPUS files encoded by the precompiled windows-binaries from this forum/thread, version 1.3 (Beta and Release Candidate) seem to be incompatible with Foobar2000 1.3 und 1.4 Beta. I always get the message "Unable to open item for playback (Unsupported format or corrupted file)". Chrome and Android can playback the same files withot errors.

Foobar2000 can playback OPUS-files from YouTube and created by the binaries of OPUS 1.2, downloaded from the OPUS homepage.

Now I wonder if it is a Foobar-, binaries- or OPUS 1.3 problem.

Re: Opus 1.3-rc

Reply #62
OPUS files encoded by the precompiled windows-binaries from this forum/thread, version 1.3 (Beta and Release Candidate) seem to be incompatible with Foobar2000 1.3 und 1.4 Beta. I always get the message "Unable to open item for playback (Unsupported format or corrupted file)". Chrome and Android can playback the same files withot errors.

I don't see this problem. How are you encoding the files? Maybe you can upload a short sample that supposedly doesn't play in fb2k?

Have you tried a clean install of foobar2000 (for example portable mode in a new folder)?

Re: Opus 1.3-rc

Reply #63
@Hanseat - No issues here either with playing tracks encoded with the latest RC of Opus in foobar. As Deathcrow said, try with a clean install of foobar2000 and take it from there.

Re: Opus 1.3-rc

Reply #64
I've been playing around with the --music and --speech switches for music encoding at 24 kbit/s. I have to say, --music (which seems to result in the same file as not specifying anything at that bitrate) sounds terrible. The stereo imaging and spectrum may be better than in SILK mode, but the artifacts are incredibly annoying and totally not worth it (snare drum hits are nothing but artifacts). Encoding with --speech sounds loads better to my tin ears, almost (but not quite) better in every way. Certainly so much better that I'd want it to be the default at that bitrate.

Am I wrong? Maybe my hearing is completely broken/abnormal?

I've attached two short samples for demonstration purposes.

Re: Opus 1.3-rc

Reply #65
Maybe --speech sounds better because the stereo image is narrower and more bits can be used on the stuff in the middle?


Re: Opus 1.3-rc

Reply #67
Maybe --speech sounds better because the stereo image is narrower and more bits can be used on the stuff in the middle?

That may well be true. I realize that we are losing some fidelity, but the clashing and clanging artifacts with the --music setting make it borderline unbearable to listen to (for me). The sample with the --speech setting doesn't make me want to rip off my headphones in agony.

Maybe  ;)

Try --set-ctl-int 4004=1104. Do you like the output?

What is that supposed to do? Sounds just as bad as the --music setting at 24 kbps and I can't honestly tell the difference... but I'm not sure what I'm listening for. If it has to do with the frequency range / bandwidth: My hearing tops out around 15-16khz.


Re: Opus 1.3-rc

Reply #69
12-20 kHz frequency range is important for noisy speech and music (stereo material, etc..) http://oh3tr.fi/~oh3gdd/Publications/%5B12%5D_Dallas_ICASSP2010_Ramo_Voice%20Quality%20Evaluation%20of%20Various%20Codecs.pdf
The study you cited doesn't provide support for your claim, it contradicts it. For mono speech, even though half their samples were noisy, their graph says that the difference between 10kHz bandwidth and fullband failed in their tests to be statistically significant, as did the difference between 12kHz and fullband for stereo speech. They state the same conclusion in their text: "Widening the audio bandwidth further from superwideband to fullband has less impact and the perceived quality does not improve significantly." Note also that the paper's "wideband" means 7kHz cutoff, so the difference between Opus WB and SWB/FB is less than what they indicate.

Even if FB costs less than 1kbps more than SWB, one might imagine that for speech at low bit rates the lower frequencies can use every single bit they can get while the >12kHz frequencies are essentially inaudible.

I was under the impression that the 1.0.x mode change thresholds etc had been based on plenty of listening data e.g. the 2011 nokia test. I can appreciate that those tests are out of date given the bugfixes and improvements since then, and I don't claim to be the superior expert here. But to an outsider glancing at commit logs and patches it seems like the relentless dropping of the bandwidth and stereo threshholds with each release doesn't entirely correlate to "hey we made the encoder so much better that we need dramatically fewer bits for the lower frequencies and now have bits to spare for more bandwidth." It seems to reflect shifting priorities, and it's natural to wonder why the priorities have shifted. Most charitably one could figure "they must have lots and lots of secret double blind listening test data." Least charitably one could figure "they saw that video that compared opus and xhe-aac via spectrographs and wanted to cater to the 'I want my prettier spectrographs' crowd." I'm fairly certain neither is correct.

Re: Opus 1.3-rc

Reply #70

Then something is really wrong at your end.

Okay since you're not elaborating I'm just going to assume it's a lowpass: Not sure why you'd expect me to immediately pick up on that considering all the loud and annoying noises/distortions going on.

Could you be a little less cryptic and a little more helpful? Did you listen to the two samples I provided? Do you think the "--music" sample sounds better than the "--speech" sample, even though the former is laden with loud artifacts and distortions? I'd probably prefer a narrowband transmission of that track over the reproduction offered by the music setting. I'm totally going to shut up if someone told me they'd want to listen to the --music sample over the --speech one

Re: Opus 1.3-rc

Reply #71
The Music sample at 24 kbit/s unsurprisingly sounds bad for a spacious, dynamic sound. The Speech sample is essentially slightly panned mono, and sounds very similar to a mono downmix encode (avg 26 kbit), without any special switches, except for few notes that are panned left at the beginning. The decay of the snare is slightly better, fuzzier with Speech, but not by much. My simple conclusion would be that the bitrate is not high enough for stereo sound, and I would prefer and use mono. Stereo becomes passable at 40 kbit/s or greater, and I'd only choose it at 48.

FhG AAC SBR/PS at 24 kbit sounds better on this sample, and maintains believable, smooth stereo. I'd choose it over 40 kbit Opus.

Re: Opus 1.3-rc

Reply #72
Okay since you're not elaborating I'm just going to assume it's a lowpass: Not sure why you'd expect me to immediately pick up on that considering all the loud and annoying noises/distortions going on.

It limits max bandwidth to SWB (12KHz). The lack of high frequencies alone should have made it easy for you to spot the difference.

Did you listen to the two samples I provided? Do you think the "--music" sample sounds better than the "--speech" sample, even though the former is laden with loud artifacts and distortions?

Defining better is hard with such low bitrates. Both are not good by any measure. But to my ears, music sounds better than speech, yes.

I'd probably prefer a narrowband transmission of that track over the reproduction offered by the music setting.

NB is excessive. But try with 1103 instead for 1104 for WB. It sounds terrible here. But maybe for you it sounds better!

Re: Opus 1.3-rc

Reply #73
Foobar2000 works fine. I just messed up the extensions.
Now I see that YouTube creates webm-files and opusenc does not.
FB2K does not like it when I call the files created by opusenc *.webm.
Chrome and Android do not seem to care.

Re: Opus 1.3-rc

Reply #74
Change the file name extension to out.opus and it plays. If you need webm, mux the opus file with MkvToolNix.

 
SimplePortal 1.0.0 RC1 © 2008-2018