Opus 1.5 Audio Codec Able To Make Extensive Use Of Machine Learning

[Anux]

Yes as I explained above those were not german soldiers, different language, different flag, different nationality.

Edit: Atleast your video has opus audio so we stay somewhat on topic

[murlakatamenka]

Already in Arch repos:

https://gitlab.archlinux.org/archlinux/packaging/packages/opus/-/commit/d8d013321dc3ed825c8d658c13136459be6123b1

Gotta self-compile with -march=native, codecs usually benefit from this the most, at least FLAC surely did.

[jmvalin]

Gotta self-compile with -march=native, codecs usually benefit from this the most, at least FLAC surely did.

Even without -march=native, a normal Opus build on Arm or x86 will have run-time CPU detection to auto-enable SSE4.1/AVX2/Neon/... So Opus can take advantage of newer instructions without sacrificing CPU compatibility.

[rmfx]

I find the source code of this codec absolutely unreadable, but the results are here, so, go team!

I admit that I struggle to really find the enhancement make any difference in their audio demos, but it was already good, so we can't go wrong with any enhancement.

[justinkb]

Only on Phoronix could a thread about an audio codec derail into a discussion about wokeness. These forums are tragic

[binarybanana]

Nazi germany was surprisingly diverse and colorful. Perhaps most interesting was the Bosnian division of the WaffenSS. Just imagine 20k Muslims in SS uniforms fighting commies. Yes, it was a huge shit show.

Also, yes, opus is great, getting better and better.

[Old Grouch]

Only on Phoronix could a thread about an audio codec derail into a discussion about wokeness. These forums are tragic

It's all part of the rich tapestry of life. Would you prefer they are comic? 🎭
It's not a huge derail: someone doesn't like the use of machine learning, and cites a well-known example that appears to show machine learning giving plainly wrong answers. Someone else points out the 'plainly wrong' answer is not as wrong as you might think. There were military forces fighting on the German side in WWII that were non-Aryan (using the Nazi definition of 'Aryan'). Some examples:

https://en.wikipedia.org/wiki/Indian_Legion

https://en.wikipedia.org/wiki/Kaminski_Brigade

https://en.wikipedia.org/wiki/Azerbaijani_SS_volunteer_formations#Ostt%C3%BCrkis che_Waffen-Verb%C3%A4nde_der_SS
https://en.wikipedia.org/wiki/14th_W...(1st_Galician)

This is not to say that someone asking for a 'typical' German soldier of WWII should necessarily get a depiction of a Sikh from the Indian Legion for 'diversity' and 'balance' reasons. There is a small amount of nuance here in needing to understand that there do exist exceptions: but presenting exceptions as 'typical' is incorrect,

From the point of view of reconstructing audio, using ML is not a bad idea, but has limitations, as people who designed voice codecs found out when their model was developed on 'Western' voices and produced worse-quality results reproducing Chinese.

IEEE Explore: Study the Voice QoE for Speech Codec in Chinese Environment

Perceptual Evaluation of Speech Quality Measurement on Speex Codec VoIP with Tonal Language Thai

When you move from a generic data compressor to a 'perceptual' codec, you end up making assumptions about what the output will look (sound) like. A perceptual codec that has used machine learning to construct its model will inevitably be biased by the content of its training set. This can lead to efficient coding and reconstruction of data that resembles the training set, but data from outside that set can end up being handled badly. There is a trade-off here that not all will be happy with.

[whatever]

D'oh. Assuming this is any good, it's slightly annoying when you already re-ripped most of your CD collection into Opus a few month back.

You should rip to a lossless format like FLAC and use that as source for transcodes to lossy formats like Opus. This way you only have to do the ripping once.

[murlakatamenka]

Even without -march=native, a normal Opus build on Arm or x86 will have run-time CPU detection to auto-enable SSE4.1/AVX2/Neon/... So Opus can take advantage of newer instructions without sacrificing CPU compatibility.

That's good 👍

I don't know why people talk about Nazi's here while the man behind Opus himself comes to the thread, so let me take an opportunity to thank you for your contribution to make the world a better place. I've been using Opus-powered mumble VoIP for many years now, to talk to my friends, so your tech makes people connect to each other. Thanks again!