If you’ve got working ears, and even more so if you’re interested in watching the Narrascope 2024 videos, please participate in this poll! It won’t take long and comparing audio excerpts is fun.
Re: Narrascope 2024 videos, you’re probably already aware that @mifga has been busy applying noise reduction to improve the poor source audio and uploading the videos to Youtube. I volunteered to visit the source audio and see if I could do more.
I worked on a test video (@Zed 's What The Glulx) the other night, and to my taste, made Zed clearer. The side effects are
There’s more high frequency hiss, but not affecting intelligibility. In fact it’s allowing that hiss that’s allowing air back into the speech. And
The other speaker in the next room is clearer, too
I’ve used a different approach than Matt’s, avoiding heavy noise reduction in favour of other techniques (split the signal into mid and side components, treat each separately, eq, dynamic eq, light dehiss, de-ess, light compression, limit).
Before I spend any time looking to apply my approach to other videos, I want to check whether the result is actually one the majority of folks would prefer to both (a) the original recording and (b) the noise-reduced alt version currently in the Youtube videos.
I’ve got an 8-second portion of Zed’s audio here. You should listen to each version of it (these are links to tiny mp3s in a public dropbox) -
Then vote for the one you think sounds most legible or that you would most like to listen to for the whole talk. Check them against other all you like if you’re not sure at first. Thanks much.
PS No, I am not remotely offended if you don’t vote for my version. The goal of this is empirical.
Many thanks! I voted for your version because the voice is crisper. But there is an underlying low-level hiss. But overall I still find that easier to listen to than either the original or the YouTube version. Many thanks folks working on this!
I like your version; the transients on the plosives and sibilants really pop. To my ears, though, there’s a faint comb filter effect which is a little distracting. Maybe something like your M/S processor is introducing phase issues? (Or it could be from the original crew forgetting to add compensation delay to microphones at various distances.)
Did you try using an expander to de-emphasize the voice from the adjacent room? It would be less noticeable than the aggressive gating the YouTube version has on it.
It could be one or both. But certainly after I’ve treated both sides separately with plugins, the different EQs are probably twisting the phase, and I can’t quantify what Apple’s Sound Isolation effect is doing to the phases
I should retest doing the EQs in linear phase. I checked that option on the way through and found on the mid stream that the pre/post ringing instantly brought out a bunch of low rumble. Sometimes with linear phase I can’t hear a difference, so since I could, I instantly flicked it off again.
I do have an expander on the mid stream. I was flicking it on and off trying to work out if it was helping, and I decided it was just a teeny tiny wincy bit. Unfortunately on the side stream (which has the bulk of the voice frequencies) I couldn’t make one work.
On the side stream, the bulk of noise is about 28 db below the speaker’s voice. On the mid stream, it’s only 12 db below. This is why I wondered for a moment if I could just throw the mid stream away, then phase issues would go, too, but without it, the speaker’s voice is probably too denuded of bass.
I also experimented with a gate, but it variously sounded ridiculous and didn’t work, due to the terrible signal to noise ratio.
You know your stuff @FLACRabbit ! Are you able to give it a go?
I wouldn’t worry about it too much, given that this is a volunteer effort. Most people probably won’t even notice the phase cancellation if they haven’t trained themselves to listen for it.
The version you have is really quite good, and I don’t think I could do any better with the time and tools I have available. Thank you!
Are all the recordings similar or are you going to have to create a new effect chain for each talk?
I haven’t vetted others myself yet but I asked @mifga on the Discord about this and it’s possible I can use this effects chain with minor tweaks for other talks in the theatre. There are two other rooms, with what I hear are much worse sources. There’s the noise-reduced version of every theatre talk on youtube now, but only some from one of the other rooms.
Comment on the poll progress: I will wait for more data, but at the moment, 1 in 3 prefer the noise-reduced version already on Youtube. I anticipated people might because of the hiss in my version which is why I felt I should run the poll. In this context, 1 in 3 is significant and I would not consider going on to replace the noise-reduced audio in the Youtube videos of the theatre talks with my versions.
While continuing to let the poll run, what I can do is see what the audio from talks in the other two rooms is like. I hear (not literally - I haven’t listened to it yet) that it is considerably worse, in which case it could possibly be made broadly better…
The Altclean works clearly best for me … but a confounder here is, I may have only understood the words for that one fully because it was my 3rd time through listening!
Yeah. It’s just practical that I only share a sliver of audio. I think when we consider most live speaking, we rarely get 100% of the exact words, but we bridge the little gaps anyway, sometimes via context or in retrospect or by what comes after. In this topic, you’ve only been given a sliver in isolation, so there aren’t the usual helps.
Unfortunately the hiss goes hand in hand with the legibility. Reducing the hiss takes it back towards the same quality as the noise-reduced version. There’s no side-effect-free version we’ll ever have with audio this poor. It’s just a matter of deciding which side-effects or outcomes we find most listenable and/or least intolerable.
