BeyondATC Voice Tier

[ShadowN7 12]

5 hours ago, Bobsk8 said:

How many voices do the offline voices have?

150 limited to US and UK accents per what the dev just said on their Discord.  He said they have access to more accents but were not happy with how they sounded right now so are just sticking with those two for now.

Edited February 20 by ShadowN7
More details

[Bobsk8 12,750]

1 minute ago, ShadowN7 said:

150 limited to US and UK accents per what the dev just said on their Discord.

Wow, that is pretty good. 

[F1Fan 121]

30 minutes ago, ConairMSFS said:

Where did you hear that? With the base package, you already have the full working software with basic voices, as far as i know. I see the premium voices as a plus, but personally I dont really care if I'll listen to many different perfect local accents, or just a few. Basic voices (with no more payments needed) looks like they sound already much better than microsoft david or whatever ones. Using credits only for push to talk, they should last a year or more at least.

From my understanding is.............The base package is $60 which includes some credits for Premium and Basic voices to get you started.  Once you run out and used up all those credits, if you want to continue using the Premium or Basic voices there is an option to buy $$$ more credits. or, you can continue and don't have to buy any credits and use their inferior less quality voices that is included in their Live Service which I believe it is called "Off-Line" voices?...I am not positive about this though regarding the "offline" mode  This PTT business is confusing me, so it is better to get the answers from them on their Discord.

[orchestra_nl 935]

15 minutes ago, ShadowN7 said:

He said they have access to more accents but were not happy with how they sounded right now

I can relate to that. Often when AI tries to emulate a non-English speaker speaking English it sounds as it the person isn't even trying to speak English but just pronounces the English words as if they were Dutch. In reality most people will at least try to speak with an English accent, with varying degrees of succes.

That will probably change when AI gets more mature and possibly the more expensiver AI solutions will be already better at this.

Edited February 20 by orchestra_nl

[mspencer 432]

36 minutes ago, Fiorentoni said:

I just checked on discord, it's NO AI driven chat-gpt ATC (unlike SayIntentions, which indeed uses ChatGPT for the actual interactions "text"). It's just plain old scripted and coded ATC for BeyondATC, no machine learning of any kind.

(also for @caneman)

I'd rather have that tbh.

I don't trust generative AI with a complex ruleset because as soon as you start exposing LLMs to increasingly complex instruction sets and expecting them to make "decisions" isn't going to work - an LLM is a "dumb" AI model - it can interpret and give you results but it is inherently dependent on what you feed it. 

There's a video from someone flying with SayIntentions when he says something unexpected and it causes the LLM to begin hallucinating and saying not... crazy stuff... but instructions or transmissions that wouldn't be given. Typical ChatGPT stuff.

It brings me back to the early days of the publicly accessible LLM when you had people posting ChatGPT "explaining" aviation concepts to them on this very forum. Most responses were broadly accurate but once you push ever so slightly into the details or ask it slightly more detailed questions, it all falls apart and you start seeing inconsistencies, incongruities, and simply fabricated information. 

AI just isn't mature enough of a technology to accommodate what it would have to. Impressive? Yes. I'm working with a business partner who's using AI to track engagements on a marketing campaign, and it generates him a new prospect list every week that he then... uses AI to hit with email campaigns. But, that's very basic sales development rep skills that he's using it for - the current available models are very good at that, they're being given a very simple data set, engagements, points that you can map, and coming up with assumptions about them. That wouldn't be a satisfying ATC model, however.

[Tuskin38 6,532]

1 hour ago, F1Fan said:

Whatever, it is still a "Live Service" microtransaction business either way you look at it. He's going to want to entice YOU to pay for the Premium Voices or at least try...I am not knocking it, he's very smart.  He sees what makes $$$$.

The devs of BeyondATC do not get money from the user buying voice time, it all goes straight to the provider they use for generation.

Edited February 20 by Tuskin38

[Kaboki 437]

I dunno, but when I talk to my «tourguide»(a feature in Sayintentions AI that is mimicking a passenger acting as tour guide) and ask him what is that river 3 miles at 1 o’clock and he tells me, and i check the map and its correct, I find that bloody amazing.

Nothing is perfect, but this chat gpt 4 thing beat any old speak recognition ATC ever made. Also its not just chat gpt 4, its linked with other ai solutions working together. Thats why this thing cost so much. 

Edited February 20 by Kaboki

[bendead 1,085]

32 minutes ago, F1Fan said:

From my understanding is.............The base package is $60 which includes some credits for Premium and Basic voices to get you started.  Once you run out and used up all those credits, if you want to continue using the Premium or Basic voices there is an option to buy $$$ more credits. or, you can continue and don't have to buy any credits and use their inferior less quality voices that is included in their Live Service which I believe it is called "Off-Line" voices?...I am not positive about this though regarding the "offline" mode  This PTT business is confusing me, so it is better to get the answers from them on their Discord.

 

Edited February 20 by bendead

[ConairMSFS 28]

41 minutes ago, F1Fan said:

I am not positive about this though regarding the "offline" mode  This PTT business is confusing me, so it is better to get the answers from them on their Discord.

Had some troubles understanding this part too. As wrote in post previus this, now it's more clear. Still uses credits, but such a small amount that should last quite a lot.

About voices probably I'ts a matter of personal preference. For me it's fine with "offline" ones too. What I'm more curious about, will be the ability of batc to understand english with terrible pizza spaghetti accent like mine 🤣 I tried p2atc (if i remember the name correctly, may be another one) time ago, but speech recognition wasnt happy about my speech at all.

[Mike_CFII_MEL 789]

3 hours ago, Tuskin38 said:

Because they've already said where all the money goes.

Plus the program already comes with a free offline voice option. So whatever their reason for not using SAPI, it has nothing to do with money.

Sorry, but I don't buy the fact that any company in business today is only looking to break even month after month after selling the original product! Where on their product page does it state 100% of the cash generated for premium voice time goes to the providers? Don't want some statement from a mod over on discord. This statement on their web page would be big to vindicate the price point and CLEARLY show where that premium cash goes. So why not have that statement on the page, it's to their benefit?

If this company goes under how do we purchase premium voice sets going forward? Since they don't provide the service, we should be able to go to the source, how would that be done if possible. Who's the provider? Can we deal with them directly? 

Edited February 20 by Mike_CFII_MEL

[Tuskin38 6,532]

Oh it says in the very first post of this thread why they didn't use SAPI, they don't think it sounds as good as offline voice they're already using.

Edited February 20 by Tuskin38

[Drumcode 1,893]

Not a bad price point for an innovative product. Sign me up with the premium package. 

[pete_auau 1,163]

11 hours ago, caneman said:

We’ll see how the product holds up, but preliminary videos show a much better ATC AI and basic voices that are better and less robotic than P2ATC.

My thoughts, I get yours may differ being a fan of P2ATC.

 

you can download  voice  packs free  that  sounds  just as good for p2a

Edited February 21 by pete_auau

[virtuali 4,421]

The problem with AI

AI, as of today, is both incredibly interesting and powerful, but it's also too expensive, especially in the context of a flight simulator, if you want to do some kind of add-on that needs to be perceived to be fairly priced.

If you followed recent AI developments from a developer point of view, you might know that most of the libraries it works with are based on Python so, a long while ago, we thought we could do something about it, considering we have a Python engine to run our software, so it might have been a good idea to integrate it in our products. In fact, our Python engine pre-dates even the first GSX, since it was built to make a way earlier a less known product called XPOI, which was a good fit for it, since it also was a mashup of internet-based "big-data", which Python was very suited for. But that worked for us, because we could sell it for a fair price, since the data we used was free to use (Geonames and Wikipedia).

For a while, we toyed the idea of doing some kind of spiritual successor to it, but with a vastly different 2020s technology behind, that is using AI for speech recognition (no user interface, you just talk to it), AI to do replies and provide info about the surrounding world and AI voices to make it sound real.

Well, it wasn't just an idea: we have a working prototype, which was made fairly quickly, because "most of the work" is done remotely.

And here's the problem started because, once you hit the hard math behind the costs, you realize it's just not doable, at least not with the quality you wanted to achieve, because:

- Using the outdated and way cheaper ChatGPT 3.5 engine, the quality of the replies wasn't accurate enough to be taken seriously, it was just too generic, and failed to identify where you were in the world by large distances (like being off by 50nm or more) and even the quality of the replies wasn't there, even without taking into account the lower accuracy.

- With the ChatGPT 4.x engine, the quality of the replies was way better, still not perfect, but at least acceptable for something not completely "serious" like a "VFR companion". I don't think it's ready for more precise tasks like being a full blown IFR ATC, but I guess this might improve in the future, possibly with lots of extra training with custom data. The "problem" is the ChatGPT 4 engine is like 10x more expensive than 3.5.  With still a fairly lightweight use to do some tests locally, using my own OpenAI API keys, I burned 10$ of credits in just a couple of days so, clearly, unless you have users prepared to spend like 50$/month, I can't see how it could possibly used regularly in all your flights (so you could say "it was worth it")

Of course, we investigated Open-Source alternatives and, they are divided into two main families:

1) The online ones, which are free but, have a queue of requests so, the more users there are on the service, the more you need to wait in a queue to get your reply. Clearly, if you are flying over some landscape and ask your companion what's around you, it's not useful and quite frustrating, if the reply comes after 30-40 seconds, so the landmark/poi is already behind you. More importantly, you can't really predict the delay.

2) The offline ones, which use your GPU to run the AI language model, and have different system requirements, depending on how large they are. The problem with this approach is that you are also running a Flight sim, it's not like you are generating an image with AI and it's ok if you can't do much more with your PC while it comes out and, unfortunately, the GPU VRAM requirements for models that *start* to approach (they are not at that level, yet) GPT-4 quality, are more than what the whole MSFS needs, that is 8GB of GPU memory just for the AI language model and yes, we tried that, and it more or less worked only with 24GB cards. Using lighter models dramatically decreased the quality of the replies.

So, as far as language models are concerned, the main issues are:

- Good AI is just too expensive, and no user will ever accept that kind of recurring cost.

- Free online AIs is just too limited by the long queues to be really usable.

- Free offline AI just can't run together with the sim, except on a very few cards (3090s and 4090s, basically), severely limiting their potential user base.

And this was JUST the part related to the AI language model. We still have to add AI-based speech recognition, and AI-generated voices! And here, the problems are similar: while speech recognition is easier and there are alternatives, state of the art AI voices that don't sound robotic are expensive and, like anything in AI, the costs are based on usage.

So, even if you decide to NOT "make any money" on the AI part, and just pass the cost of AI to the user, it will still cost too much to be considered acceptable for the vast majority of users, and you still need to sell some product because, regardless of the fact "AI is doing most of the work", you STILL need to work on the integration and to support the product so, clearly, there must be some initial payment. That, or you decide to make it "free" and add some kind of mark-up on the AI costs.

So, in the end, we decided it was a nice learning experiment, which we'll surely want to revisit when the time is right, perhaps AI costs will drop in the future due to more powerful hardware, or we might see the appearance of consumer-level GPUs with dedicated AI chips and memory, so you can run your game and the AI together, locally, with no performance loss.

So, our XPOI AI-based successor project has been put on hold and, funny enough, several months after we abandoned it, somebody had a similar idea and announced it recently, it's called AviMate. I wish them luck.

Edited February 21 by virtuali

[F737MAX 5,616]

4 minutes ago, virtuali said:

The problem with AI

AI, as of today, is both incredibly interesting and powerful, but it's also too expensiv

Did AI write all that? 😆

 

Fully agree that good AI is too expensive, free is too limited and offline too intensive.

Plus, AI doesn't always do what you want it to. Apparently, SayIntentions.ai can get lost in a ChatGPT style conversation, which is exactly what you don't want for an ATC product.