Podcast Ep. 01: Scenema Audio, Open Source vs Audio Pro

In this first episode, Grace and Jude pull apart Scenema Audio’s release strategy: a fully open source 22B-parameter generative audio engine on one side, a paid Studio Pro tier on the other, and a genuinely counterintuitive product decision sitting between them. The conversation covers the actual model stack (LTX 2.3 extraction, Gemma 3 encoder, SeedVC, MelBandRoformer, Kokoro), the hardware bar for running it locally, the workflow friction that separates infrastructure from production, and the philosophy of addition by subtraction that led the team to remove sound effect generation from the paid product.

Transcript is diarized from the source audio and lightly cleaned for readability. Model and product names have been corrected.

Transcript

Grace. Imagine for a second that you walk into this high-end car dealership. You are in the market for a really sleek, finished sports car.

Jude. Right. You want the full package.

Grace. Exactly. But instead of handing you the keys to a vehicle, the salesperson just points to a wooden pallet sitting right in the middle of the showroom floor.

Jude. Oh, no.

Grace. Yeah. And resting on that pallet is this massive, pristine, unbelievably powerful engine. And the salesperson looks at you and says, hey, the engine is completely free. You can literally take it home right now.

Jude. Okay. But what’s the catch?

Grace. Right. They say if you actually want the steering wheel, you know, the seats, the chassis to drive it, well, you have to pay a monthly subscription.

Jude. Wow. Which, I mean, it feels like a trap, doesn’t it?

Grace. It really does.

Jude. Because most companies guard their core algorithms like state secrets. Giving away the absolute most expensive piece of technology for free seems, well, entirely counterintuitive to actually making money.

Grace. Which is exactly why we’re pulling from a massive stack of developer notes and GitHub repositories for this deep dive today. We are unpacking the really fascinating product strategy of a new AI startup called Scenema Audio.

Jude. Yeah, we should probably clarify something right off the bat for everyone listening.

Grace. Oh, yeah, the pronunciation. So in the source code, it is spelled S-C-E-N-E-M-A. But the team pronounces it exactly like their word cinema.

Jude. Right, which is a great bit of wordplay. I mean, it hints at their ambition. They are clearly aiming at the cinematic professional audio market.

Grace. Yeah, and Scenema has dropped this incredible AI generative audio model. And they did exactly what we just described with that sports car analogy. They took their core model, this incredibly powerful engine, and they just released it completely open source.

Jude. Which is a huge deal.

Grace. It is. Anyone can go to GitHub or Hugging Face right now, download it, and run it for free. Which, you know, immediately begs the question, if the engine is free, why would anyone ever pay for their premium tier, which they call Studio Pro?

Jude. Well, to figure out if they’ve completely cannibalized their own business, we first have to look at what they’re actually giving away. Because this open release is not just, you know, a single simple algorithm.

Grace. Right, it’s a lot more complex than that.

Jude. It’s a brilliant case study in how modern AI is actually built. Looking through the repository nodes, you realize this isn’t one monolithic brain. It is an ensemble.

Grace. It’s like a Frankenstein’s monster of different specialized AI models, right? All stitched together to produce audio.

Jude. Exactly.

Grace. So if I’m going to drop real money on this pro tier, I need to know exactly what I’m competing against for free. Let’s look under the hood of this open source engine. What are the actual nuts and bolts here?

Jude. So the heart of the system is an audio diffusion transformer. And the development team actually extracted this from a model called LTX 2.3.

Grace. Okay, LTX 2.3.

Jude. Yeah, which is this massive 22 billion parameter audio visual model. It was originally designed to understand both sight and sound simultaneously.

Grace. So instead of training a completely new audio generator from scratch, Scenema essentially surgically extracted just the vocal cords from a model that already had a massive understanding of how the real world sounds.

Jude. That surgical extraction is exactly what makes this so unique. They isolated the audio generation capability. But, you know, an audio generator is completely useless if it doesn’t understand what you actually want it to make.

Grace. Right, it needs a brain to process your instructions.

Jude. Precisely. So they paired that LTX extraction with the Gemma 3 12B text encoder.

Grace. And Gemma 3 is a large language model. So its job in this whole pipeline is to read the user’s text prompt, parse the context, and translate our human language into like a mathematical map that the audio generator can actually execute.

Jude. And it does that heavy lifting brilliantly. Though just a quick technical note for anyone listening who wants to pull this from GitHub.

Grace. Oh, the token thing?

Jude. Yeah. Because Gemma 3 is a gated model, you do need to supply a hugging face token in the configuration file to authorize the download. It won’t just work automatically.

Grace. Good to know. Okay, so we have the brain, and we have the vocal cords. But the sources highlight some highly specialized organs working alongside them, right? Starting with something called SeedVC.

Jude. Yeah, SeedVC is amazing. It handles voice identity transfer. In plain English, that just means zero-shot voice cloning.

Grace. Zero-shot, meaning you don’t have to train it for hours.

Jude. Exactly. You feed the system a really short reference clip, literally just 8 to 16 seconds of someone speaking, and SeedVC analyzes the unique acoustic fingerprint of that person.

Grace. That’s wild. Just 8 seconds.

Jude. Yeah, it maps the timbre, the resonance of their vocal tract, their specific cadence, and that allows the engine to replicate that voice perfectly on entirely new text.

Grace. But, okay, capturing a perfectly clean 8-second clip in the real world is pretty tough. Like, if I record my reference sample on my phone on a busy street, I really don’t want my AI clone to constantly have traffic noise playing underneath my voice every time it speaks.

Jude. Which brings us to the next organ in the ensemble, which is called MelBandRoformer.

Grace. MelBandRoformer. That is a mouthful.

Jude. It is a bit of a mouthful, but its sole purpose is vocal separation. And it doesn’t just act like a dumb volume knob for background noise. It analyzes the spectrogram of the audio.

Grace. So, like, the visual representation of the sound waves?

Jude. Right. It identifies the specific frequency bands that belong strictly to human vocal cords, and it mathematically subtracts everything else.

Grace. Oh, wow.

Jude. Yeah, whether that’s a car horn, wind, or even room echo. It strips it all away and feeds SeedVC a pristine, isolated vocal track.

Grace. That is incredibly elegant. So you’ve got the language parsed, the voice cloned, the background noise totally stripped out. But there’s one more piece mentioned in the white papers. It’s a much smaller model called Kokoro, sitting at just 82 million parameters.

Jude. Right, Kokoro.

Grace. The sources say it handles duration estimation. How does an AI actually estimate duration?

Jude. Think of Kokoro as the orchestral conductor of this whole ensemble. Because if you want synthetic speech to actually sound human, pacing is everything. You can’t just spit out words at a uniform robotic speed.

Grace. Right, that’s the classic Siri voice from 10 years ago.

Jude. Exactly. So Kokoro works by analyzing the statistical rhythm of language. It looks at your text and understands that, hey, a comma requires a certain millisecond pause. Or that ending a sentence on a question mark fundamentally changes the speed of those last few syllables.

Grace. It actually assigns precise time values to every phone so the delivery flows naturally.

Jude. Yes. So all of these specialized models, Gemma, LTX, SeedVC, MelBandRoformer, and Kokoro, are all firing in sequence to produce a single audio file.

Grace. So what does that final product actually sound like? Like, what’s the quality?

Jude. The output is honestly stunning. It generates 48 kilohertz stereo WAV files. Let’s contextualize that for a second for the listener.

Grace. Yeah.

Jude. 48 kilohertz is not that compressed, slightly tinny audio you get on a Zoom call or even a standard MP3. 48 kilohertz is the uncompressed standard for professional film and television production. Yeah. It’s broadcast ready right out of the box. And the open release supports 13 different languages natively. 13. Yeah, including English, German, French, Spanish, Japanese, Arabic, and Swahili. It’s a massive global footprint.

Grace. Okay, so the engine is undeniably powerful. But how do you actually interact with it? Because looking at the deployment notes, you aren’t just, you know, clicking nice buttons in a slick, colorful app.

Jude. No, not at all. The open source version relies entirely on an XML prompt syntax.

Grace. Right. It feels very much like formatting a screenplay rather than using traditional software. You literally type out your text, but you wrap it in specific code tags.

Jude. Exactly. You might write an opening bracket, then type speak voice equals. You put a name, scene equals, language equals English, and then close the bracket.

Grace. And inside that container, you can use action tags to give the AI delivery directions, right? You can program the voice to whisper or to sound sarcastic or to project as if they are standing in a large, echoey hall.

Jude. It is literal audio programming.

Grace. But here’s the detail that really caught my eye in the notes. In this free, open version, there is no preset voice library.

Jude. None at all.

Grace. There is no drop-down menu where I can just, you know, click and select friendly corporate announcer or gruff action hero.

Jude. Nope. You have to design the voice entirely from prose descriptions.

Grace. Which means I literally have to type out a 45-year-old woman with a slight British accent who sounds like she hasn’t slept in three days.

Jude. Yes. And then you wait to see how the AI interprets that description.

Grace. It has to take a while.

Jude. Oh, it requires an immense amount of trial and error.

Grace. Yeah.

Jude. Because it’s wildly creative. But you are essentially directing an actor who doesn’t even exist yet. You’re just tweaking adjectives until the voice matches that sound you have in your head.

Grace. Which means you are going to be generating a ton of audio over and over again. And to run this entire Frankenstein operation locally on your own machine, the hardware admission ticket is incredibly steep.

Jude. Extremely steep. The weights for these models, like, the actual neural connections that make up the AI’s brain are massive. The standard BF-16 version, which uses high-precision math, is a 9.8 gigabyte download all by itself.

Grace. And BF-16, that stands for Brain Floating Point 16, right?

Jude. Right.

Grace. It’s a format that basically gives the AI highly precise numbers to work with. But as you said, it takes up a lot of space.

Jude. Correct. Now, they do offer a compressed INT8 version, which shrinks the file down to about 4.9 gigabytes.

Grace. So half the size.

Jude. Yeah. Think of INT8 like packing a suitcase really tightly. You round off some of the mathematical decimal points to save space. You lose a microscopic amount of audio precision, but it runs much faster.

Grace. Okay. But how do you actually get it running?

Jude. To deploy either version, developers run a simple terminal command docker, ComposeUp. This packages all the dependencies and spins up a local server right on your machine, creating a little endpoint at local host port 8000.

Grace. So everything is happening entirely on your own hardware. No data is going to the cloud. You have total privacy.

Jude. Total privacy, yes.

Grace. But to actually process all this, the documentation clearly states you need an NVIDIA RTX 4090 graphics card with a minimum of 16 gigabytes of VRAM, though 24 gigabytes is highly recommended.

Jude. We should definitely explain why that VRAM requirement is so crucial. VRAM is video random access memory. When you run an AI model locally, your computer has to load all 22 billion parameters of that LTX extraction into active memory all at the exact same time.

Grace. So if it doesn’t fit.

Jude. If you don’t have 24 gigabytes of space on your graphics card, the model simply won’t fit. It will try to use your system’s regular slower RAM, and the generation process will slow to an absolute crawl, or it’ll just crash entirely.

Grace. And an RTX 4090 with 24 gigs of VRAM is basically a cinder block of silicon that costs well over $1,000. It is absolute top-of-the-line hardware.

Jude. The software is free, but the hardware is definitely not.

Grace. Right. Okay, so this brings me back to my initial pushback. Let’s say I’m, I don’t know, a moderately tech-savvy YouTuber. I already own that cinder block RTX 4090 because I use it for video editing anyway.

Jude. Sure.

Grace. I can clone voices, I can design custom characters from pros, and I can generate professional 48 kilohertz audio in 13 languages, all for free. I can easily learn XML syntax in an afternoon. Isn’t Scenema losing the exact middle class of creators they want to attract? Like, why would I ever pay a monthly fee for their Studio Pro tier?

Jude. You are looking at the raw capability and assuming it translates directly into a usable daily workflow. The crucial distinction here is the difference between building infrastructure and actively producing content on a deadline.

Grace. I’m not sure I totally buy that. I mean, if the engine works, it works.

Jude. Let’s look at the editorial reality highlighted in the product white papers. The Studio Pro tier is not a fundamentally better AI model under the hood.

Grace. It’s not.

Jude. No, they aren’t hiding a smarter algorithm behind a paywall. Pro uses the exact same LTX diffusion and Kokoro duration estimation we just spent the last 10 minutes discussing.

Grace. Okay, so what am I paying for?

Jude. What Pro adds is a massive layer of training extensions and a completely frictionless production wrapper. Let’s just start with scale. The open release supports 13 languages. The Pro tier jumps to 30 languages.

Grace. Okay, well, if I’m a global brand localizing commercials, having 17 additional languages instantly justifies the cost. That’s huge.

Jude. It is, but the real value lies in endurance and time. The Pro tier offers verified, single output generation for up to 30 minutes continuously using something called automatic chunking.

Grace. Ah, automatic chunking. Let’s dig into that because understanding how context windows work is really the key to why the free version is actually a massive headache for long-form creators. AI models have a limited short-term memory, right?

Jude. Exactly. That’s the context window. If you feed a local AI model a sublimate audio book chapter all at once, the sheer amount of text and audio data will just overflow its memory buffer and your computer will crash.

Grace. So if I’m trying to be frugal and use the free open source version for a 30-minute script, I can’t just press play. I have to physically chop my script up into tiny, you know, two-minute chunks. So I paste chunk one, I hit generate, I wait, I save the file, I paste chunk two, I wait, I save it. And then I have to open a program like Audacity and manually stitch 15 different audio files together myself.

Jude. And you are praying that the AI didn’t randomly shift its emotional tone or its volume between chunk seven and chunk eight.

Grace. Does that happen a lot?

Jude. It happens constantly in raw models.

Grace. Wow, that sounds like an absolute nightmare if you are on a production deadline.

Jude. It is incredibly tedious. And that is exactly why Pro’s automatic chunking is the chassis of that sports car. Behind the scenes, the Pro software is automatically doing all that chopping, generating, analyzing for consistency, and seamless stitching for you.

Grace. It just handles it.

Jude. Yeah, you paste a 30-minute script, you hit generate, and you get one flawless 30-minute audio file. If your job is to create content, the hours of manual labor you save make the subscription pay for itself instantly.

Grace. It’s the difference between being a mechanic who loves tuning the engine in the garage and a driver who just needs to get to work on time. Pro lets you actually drive.

Jude. That sanded-down workflow extends everywhere, too. In the free version, we talked about designing voices through trial and error with Pros. In Pro, they introduce a curated system voice library.

Grace. So the drop-down menu exists in Pro?

Jude. Yes. You still have the Pros option if you want it, but you also get immediate access to a gallery of professionally-tuned, highly consistent voices that are just ready to go.

Grace. So you don’t have to reinvent the wheel every Tuesday when you have a new script. And the SourcesNote Pro also gives you a visual timeline, a voiceover inspector for microedits, and a version history to track different takes. It’s actual, usable software, not just a command-line prompt.

Jude. Exactly. They’ve totally removed the friction between having an idea and holding a finished file.

Grace. Okay. So the Pro tier is a very clear case of addition. You get more languages, longer generations, better workflow tools. But here is where this deep dive gets really interesting. Because while it’s impressive what the Pro tier adds, the most fascinating product decision Scenema made is actually what they chose to subtract from the paid version.

Jude. Honestly, this specific decision says more about the future of AI product design than anything else in the white papers.

Grace. Let’s go back to the free open-base model on GitHub for a moment. In that XML syntax we talked about earlier, there are also sound tags. In the free version, you can theoretically type sound tag dog barking or sound tag gentle breeze, and the core model will generate that environmental audio underneath your voiceover.

Jude. The raw capability for sound effects, or SFX, is built into the underlying architecture, yes.

Grace. But if you upgrade, if you actually pay a monthly fee for the premium Studio Pro tier, those sound effect tags are completely stripped out. They are gone.

Jude. They actively removed the environmental sound generation from their flagship professional product.

Grace. Which sounds completely crazy. Why would you delete a feature from the version people are paying for?

Jude. My deduction here, thinking about how those frequency bands work with MelBandRoformer, is that if the AI is trying to render complex human speech and a dog barking simultaneously, those audio frequencies are going to bleed into each other and create, like, weird digital artifacts.

Grace. Your deduction is spot on. The development team was refreshingly honest about this. They judged the base model’s sound effect generation to simply be too unpredictable for professional use.

Jude. Unpredictable how?

Grace. Well, sometimes a prompt for a gentle breeze sounded great. But other times, the frequency overlap caused it to sound like a distorted jet engine. Or a dog bark suddenly turned into this loud burst of digital static. Oh, yikes. But in a free, open-source sandbox, a developer doesn’t really care if a feature fails three times before it works, right? They are just experimenting.

Jude. Exactly. But in a professional studio environment, reliability is paramount. If a sound designer is mixing a commercial on a tight deadline, they cannot use a tool that only works 70% of the time.

Grace. It’s like a car manufacturer removing a really flashy, futuristic, autonomous driving feature from their professional trucking fleet.

Jude. Yeah, that’s a great way to put it.

Grace. Because it looks incredibly cool in a YouTube demo to say, hey, look, the truck can drive itself. But if you manage a logistics company, you don’t want a truck that drives flawlessly 90% of the time but randomly swerves into a ditch the other 10%.

Jude. No one wants that truck.

Grace. Right. You need the tool to be perfectly predictable, even if it means stripping out the futuristic bells and whistles.

Jude. That is a perfect parallel. What Scenema is practicing here is the philosophy of addition by subtraction. In the tech industry right now, there is an overwhelming temptation to leave every single experimental feature turned on, just so the marketing department can have the longest possible feature list on their website.

Grace. Right. They want to say, look at all these amazing tags our AI supports. We literally do everything.

Jude. But by deliberately removing the unpredictable SFX features, Scenema is actively protecting the professional user from the technology’s current limitations. They are basically telling their paying customers, we only want you to rely on the parts of this model that are rock solid.

Grace. That actually builds immense trust. It shows they truly understand the realities of a production environment where a predictable tool is infinitely more valuable than a novel one.

Jude. It proves they are building a product for real professionals, not just a flashy tech demo for investors.

Grace. So, synthesizing everything from these white papers and the GitHub repositories, what does this mean for you, the person listening right now? How do you actually choose which version of Scenema Audio to engage with?

Jude. The practical guidance ultimately comes down to what you are trying to accomplish and, frankly, what kind of friction you are willing to tolerate.

Grace. Let’s start with the developers.

Jude. So, if you are a developer or just a tinkerer who happens to have that massive RTX 4090 sitting on your desk and you want to build a custom application on top of a powerful engine, like maybe you are coding a video game and you want to automate thousands of dynamic short character barks.

Grace. Yeah.

Jude. The open release is genuinely capable. You should self-host it, dive into the XML, experiment with the prompt weights, and just use it as infrastructure.

Grace. But if you are a creator, if your core job is making long-form YouTube essays or producing audiobooks or managing localized marketing campaigns…

Jude. Then the Studio Pro tier is built for you. Your time is far too valuable to spend manually chunking text files, stitching WAVs in Audacity, and troubleshooting Docker containers all afternoon.

Grace. Right. You need that sanded-down experience.

Jude. You need the chassis and the steering wheel so you can focus entirely on creating rather than managing the engine.

Grace. It really is a masterclass in modern software strategy. By giving the core engine away for free, Scenema gets thousands of developers battle-testing their code, improving the architecture, and proving to the world that their technology is world-class. It is total transparency. But they know that the people who actually produce content for a living will happily pay for the interface that gives them their time back.

Jude. It is the ultimate show-don’t-tell.

Grace. And if you want to look at the engine on the pallet for yourself, you can go directly to the source code right now. You can find it at github.com/ScenemaAI/scenema-audio.

Jude. I highly recommend reading through their documentation, too. Even if you never deploy the code yourself, understanding how models like MelBandRoformer and Kokoro actually fit together provides a brilliant window into the nuts and bolts of modern generative AI.

Grace. It totally demystifies the magic trick. And it leaves me with this one lingering thought for you to chew on as we wrap up today. We are entering an era where raw, multi-billion parameter AI models are increasingly being given away completely for free. The algorithm itself is rapidly becoming a commodity. So as these open source engines become ubiquitous everywhere, is the future of the software industry going to be less about who has the smartest AI and almost entirely about who can build the most frictionless, reliable, human-centered user interface around it? Are we moving from an arms race of raw capability to an arms race of pure usability? Something to think about the next time you see an incredible piece of technology released for free and find yourself wondering, where’s the steering wheel?