[MLA] Lossy encoder clipping
Ian Shepherd
ian at mastering-media.co.uk
Sat Oct 13 22:17:57 CEST 2018
Hi Kevin,
Perhaps you’re right, but Eelco’s quote from Spotify says
> The decoder(s) is/are float, but we clamp them to int right after the decode step
Assuming that’s correct, the move to an integer value will “bake in” decode overlevels, after which they’re then applying normalisation and the user volume control setting, followed by their limiter, all inside the App. None of that is handled by the OS, to my knowledge.
Hopefully Eelco can confirm one way or another.
Ian
> On 13 Oct 2018, at 17:28, Kevin Gross <kevin.gross at avanw.com> wrote:
>
> Ian, I think you misunderstand what Eelco said about Spotify. They do have a floating-point decoder, but they apparently hand the audio off to the OS in integer format. It is in this handoff where there are potential problems. Modern OS audio can carry floating point through the system. Apparently Spotify isn't using these capabilities to full advantage. This is not at all an unusual situation. OS floating point audio support is fairly recent and if you want your app to work on any legacy platforms, you need to present integers.
>
> In any case, you do eventually have to convert to integer to feed your DAC or or digital audio interface and if the OS audio system is set up for unity gain (which most still consider best practice) there will still be problems.
>
> I hope/expect you can workaround this shortcoming in Spotify by choosing their "Quiet" loudness normalization or taking down the gain control in their desktop application by a few dB.
>
> Kevin Gross - AVA Networks
>
>
>> On Sat, Oct 13, 2018 at 8:11 AM Ian Shepherd <ian at mastering-media.co.uk> wrote:
>>
>> Correction - the limiter I mentioned is in use on Android, not iOS.
>>
>>
>>
>>> On 13 Oct 2018, at 15:04, Ian Shepherd <ian at mastering-media.co.uk> wrote:
>>>
>>>
>>> HI All,
>>>
>>> I’ve been talking to Robert more about this, and also I emailed JJ to get word from the horses mouth.
>>>
>>> It seems I miss-spoke in my blog post - mp3 and AAC encoders (and from the Spotify comments, Ogg/Vorbis also) should encode in floating-point, and store the data with what JJ refers to as “scalefactor information (kind of like a coarse floating point)”
>>>
>>> In other words, a well-written encoder won’t “bake in” the clipping (my bad). But of course, we can’t assume that all encoders are well-written :-/
>>>
>>> Robert and JJ both agree that not all decoders operate in floating-point and Eelco’s information confirms this to be true for Spotify, sadly - so this may be the source of the bad experiences we’ve all had. If a decoder resolves to fixed-point without adjusting for headroom and before normalising, the clipping may as well be "baked-in”, as far as the final result is concerned :-/
>>>
>>> (iTunes does use floating-point, but also employs a limiter to prevent clipping, apparently…! I’m going to try and find out where this limiter is placed in the chain - I bet you a cup of really hot tea that it’s right after the decoder :-/ )
>>>
>>> I’m clarifying a few details with JJ and will then update my post - the bottom line is that avoiding high peaks is still good advice, although not quite for the reasons I originally posted about.
>>>
>>> We learn new stuff every day !
>>>
>>> Ian
>>>
>>>
>>>
>>>
>>>> On 13 Oct 2018, at 12:23, Bob Katz <bobkatz at digido.com> wrote:
>>>>
>>>> Yes, Thomas, you have shown this many times and this is true. But it seems the clipping does not happen in the encoder if the encoder and decoder are floating point (as most are). The clipping actually occurs not in the decoder either, but when the output material of the decoder is converted to fixed point. Which is why Robert from Fraunhofer protested when Ian Shepherd made the contention that the clipping is "baked in".
>>>> It would not take much effort to prove this. I've got clients in all day. I'm not sure if I'll have time to demonstrate (or disprove) this contention before AES NY. Regardless, I agree with you, I would not recommend encoding material that would clip. So for me this is a purely academic exercise.
>>>>
>>>> Best wishes,
>>>>
>>>>
>>>>
>>>>
>>>> Bob
>>>>
>>>>> On 10/12/18 12:41 PM, Thomas Lund wrote:
>>>>> Companeros,
>>>>>
>>>>> With lossy encoders, clipping can happen in the M/S domain, even if e.g. the S signal is nowhere near clipping, but the M component is. I’ve played several listening examples of that on panels with you, Bob and Bob. Sample rate converters is another place where you can’t take pre-attenuation let alone headroom for granted.
>>>>>
>>>>> Better to stay below 0 dBTP, and to keep checking M/S.
>>>>>
>>>>> cheers,
>>>>> Thomas
>>>>>
>>>>>
>>>>>
>>>>>> On 12 Oct 2018, at 18.34, Bob Katz <bobkatz at digido.com> wrote:
>>>>>>
>>>>>> Thanks, Eelco, for that confidential information. You have a good contact at Spotify. Keep on pushing them to stop that egregious peak limiting!
>>>>>>
>>>>>> Even though with a floating point codec, the clipping is NOT officially "baked in", I would always advise engineers to protect the path ahead of the encoder. It's the safest and probably the best place to deal with future clipping down the line. Apple performs the encodes and adds metadata. Engineers and producers do NOT send AAC files to Apple. And you won't find Apple changing data, even if it would prevent clipping down the line. Apple would NOT attenuate a master in front of the encode. This is a mastering engineer's prerogative.
>>>>>> And the chances of expecting a playback device to inspect ahead of time for overshoot during a streaming session are slim and none. Think about it, the real time decoder would have to have foreknowledge that the codec was going to overshoot, and know ahead of time to attenuate the whole song before it had started. It's just not practical. And I don't believe in adding peak limiting to already-mastered and approved material. At the least it changes the producer's intent, at worst it distorts and just goes one step closer to sucky broadcast-style processing.
>>>>>> So I would not change my practices because of this discovery. Prevention is the best way to a cure.
>>>>>>
>>>>>>
>>>>>> Bob
>>>>>>
>>>>>>
>>>>>>
>>>>>>> On 10/12/18 12:10 PM, Grimm | Eelco wrote:
>>>>>>> Hi Kevin and Bob,
>>>>>>>
>>>>>>>> Robert from Fraunhofer was just complaining because Ian claimed that the clipping was "baked in" to the encoder. Which is untrue to the best of my knowledge. The encoder/decoder are in a closed floating point relationship and so if you intercept the decode in the floating point domain, attenuate it and dither it to 24 bits and then send it on to the DAC or file or whatever, there will be no clipping.
>>>>>>>>
>>>>>>>
>>>>>>> I got a reply from my Spotify contact. Please treat this as confidential information!
>>>>>>>
>>>>>>> "I've also found out the float/int situation, and unfortunately it was as I feared, we're using int:s throughout the pipeline. The decoder(s) is/are float, but we clamp them to int right after the decode step. This is also, for sure, on our roadmap to change, but no promises on timeline there either, I'm afraid."
>>>>>>>
>>>>>>> So at the moment there’s not much they can do. But at least they are aware of the problem and the opportunities now.
>>>>>>>
>>>>>>> All the best,
>>>>>>> Eelco
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> _______________________________________________
>>>>>>> MLA mailing list
>>>>>>> MLA at grimmaudio.nl
>>>>>>> http://grimmaudio.nl/mailman/listinfo/mla_grimmaudio.nl
>>>>>>
>>>>>> --
>>>>>>
>>>>>> If you want good sound on your album, come to
>>>>>> Bob Katz 407-831-0233 DIGITAL DOMAIN MASTERING STUDIO
>>>>>> Author: Mastering Audio
>>>>>> Digital Domain Website
>>>>>>
>>>>>> No trees were killed in the sending of this message. However a large number
>>>>>> of electrons were terribly inconvenienced.
>>>>>> _______________________________________________
>>>>>> MLA mailing list
>>>>>> MLA at grimmaudio.nl
>>>>>> http://grimmaudio.nl/mailman/listinfo/mla_grimmaudio.nl
>>>>
>>>> --
>>>>
>>>> If you want good sound on your album, come to
>>>> Bob Katz 407-831-0233 DIGITAL DOMAIN MASTERING STUDIO
>>>> Author: Mastering Audio
>>>> Digital Domain Website
>>>>
>>>> No trees were killed in the sending of this message. However a large number
>>>> of electrons were terribly inconvenienced.
>>>> _______________________________________________
>>>> MLA mailing list
>>>> MLA at grimmaudio.nl
>>>> http://grimmaudio.nl/mailman/listinfo/mla_grimmaudio.nl
>>
>> _______________________________________________
>> MLA mailing list
>> MLA at grimmaudio.nl
>> http://grimmaudio.nl/mailman/listinfo/mla_grimmaudio.nl
> _______________________________________________
> MLA mailing list
> MLA at grimmaudio.nl
> http://grimmaudio.nl/mailman/listinfo/mla_grimmaudio.nl
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://grimmaudio.nl/pipermail/mla_grimmaudio.nl/attachments/20181013/b725f3bb/attachment.html>
More information about the MLA
mailing list