Okay, I think it's better. Nemotron seems to be real time again. I think it was a configuration issue. I also went back to 608 from 708. And I also looked up the BBC broadcast standards and tried to implement them the best I could. I would love it if you guys could try my image ghcr.io/mackid1993/ah4c:beta. Let me know how badly I screwed up. 
and I'll continue to iterate.
You probably already know this, but Google's AI spit out the following for CC "best practices":
Display Duration: Minimum of 1 second for single-line text, 1.5 seconds for two lines, and 2 seconds for three lines.
Speed Limits: Generally capped between 120 to 160 words per minute depending on the target audience and medium.
And I think we can safely say, that the "target audience" here is mostly "old dudes" (not including you of course). 
I'm taking a look at this now 
I think it would be best to make it configurable as well.
If you pushed your changes to PR#20, I'm going to build from that. Right at the moment you're pushing your PR commits against "main", and I'm pushing mine to "beta", so you haven't picked up changes I've made -- which is fine for now.
I'm pushing to PR #20 since I don't have repo access. I just made another small change that might help old dudes!!
1 Like
I think it's actually a lot better now. I was just watching CNN and I could read it for sure. Although I'm 32 and I have young eyes. 
:beta3 now has all of your current commits.
Seems better. I set the reading speed at 130 wpm. I still think the minimum time on screen should be adjustable (if possible). Also, would it be possible to not have a blank line between CC lines? I think that could make reading easier as well.
(post deleted by author)
It's a totally misguided setting that should not have been there probably. It made it wait longer, which would have just thrown the transcriptions off of speech. I believe I fixed the blank line.
These are what the settings should do now. Sorry for the British English, I don't know why Claude keeps outputting that.
Actually watching this on my TV right now, it actually looks really good.
To my eyes, so I'd like to hear what other people have to say. I think that ability to slow it down should help everyone though.
For me, setting reading speed to 150 WPM and setting time on screen to 5 seconds looks fantastic.
Edit:
Okay, well, five seconds it falls behind. I would say don't go over four seconds unless you have to. That might work better with a faster model because I was testing with Cohere.
Ive been playing with the updated beta3. Ive used both the Cohere and the Nemotron models with various settings.
The presentation of both has improved, so there aren't the issues with a flurry of buffered text showing up and quickly then disappearing, like me and @bnhf observed earlier. I settled on 130 WPM and the Nemotron model for now.
However, at least on the Nemotron model, which I like a little better, Im noticing that the text delay after watching for several minutes seems to increase at times to up to 20 seconds behind. Eventually it resets and re-syncs itself.
I tried playing with the roll up speed but had no affect...
Ill keep playing more, but definitely improved since the first beta3
The problem with setting any of these settings too low is that it will cause a drift, as it's trying to keep up with the speech.
I would say try to tweak it as little as possible. Maybe try Cohere; it buffers a sentence at a time, so it should produce a bit slower.
Just tested Nemotron and I see how it's a little too fast because it's real time. I would say try 140 words per minute and 5 seconds
@mnwxman132 what kind of Mac are you running on?
I'm wondering if the transcriptions are entirely unbounded. And someone with a really fast machine, unlike my Unraid server, might be getting really, really fast transcriptions. That could be why it looks fine for me, but you guys are having issues
Edit:
I confirmed they are not bound to the speed of the computer. They try to keep pace with the actual speech. So, if you try to really push it off, it's going to fall behind speech and if you're using a real-time model, it's going to be really fast
I think the idea is to play with all the different settings and just find what works.
If you want, I can add a setting to make Nemotron not real time if you prefer it. I think that's the whole point of the model. kind of defeats the purpose
Im on an M4 Mac mini, with 16gb internal unified memory, 256gb internal memory.
What I was thinking that might make your life a lot easier, is to give those that are interested a set of baseline test program material. So that when we give you feedback with this very valuable tool that you have unveiled, it's all from the same program material - at least at this early on stage. This way you aren't getting a million different subjective opinions from a million different people using a million different video sources. Perhaps initially, you may want to scale this back for your own sanity. If you provide us with
1: a 3 or 4 minute recording of a live sports broadcast (English preferred by me)
2: a 3 or 4 minute recording of an action TV program with lots of dialog
3: a 3 or 4 minute recording of a News broadcast
4: anything else you can think of
5: All or None of the above (lol), and the length of the clip is up to you.
We (or at the very least I ) can easily figure out a way to play these back from my Apple TV or ONN box into the linkpi.
Then at least at first, we can all be using the same test material. A lot of this is subjective, but this could help narrow down some of the variables for you as we give you this feedback..
This is totally your choice, but perhaps this could help you from getting pulled in a million directions?
Keep up the fantastic work!
I think that would be really hard to pipe through ah4c.
I think people are just going to have to tweak the settings until they figure out what works for them.
What are you seeing that isn't working for you? Like is it still too fast or is it falling behind?
Because the problem is you can't really have both. If you want it to keep up with speech, it can't hang on the screen for too long. I would say try maybe four lines. Keep the words per minute close to the default, maybe a little bit slower. Increase the time on screen a little bit. 150 words per minute and 5 seconds is working well for me.
I've been watching at that pace for the last hour or so, and it's been just perfect
I will play around with the default settings again tomorrow for the different models, and then tweak the settings from there and get back to you.
Honestly im thrilled that @bnhf was able to merge the apple TV folks into the main ah4c tree and I didn’t expect to get so much up and running this quick.
Get some well deserved rest!!!
Thanks, yeah, I was up until 3 a.m. working on this yesterday, so I'm a little burnt, but hopefully, some more people will test, we'll get some feedback, and be able to refine from there.
I got a question.
any chance we can do something about the tuner numbering? as you can see I originally tried to plan my ip's logically to help keep everything strait to my eye 
and in my brain 
What do you mean? Tuner numbering is unchanged? Goes from tuner 0 to tuner whatever.