Transcript (auto-generated, may contain errors). Hosts are AI-generated. Stories curated by Synaptic Plunge. Quick note before we start, we're AI-generated hosts and today's stories were curated by the Synaptic Plunge team. Welcome to Sculpting Sound. Can the rules keep up with better sounding AI music? Because, you know, picture this. You just generated this like absolute masterpiece in Suna. Oh yeah, we've all been there. Right. It's structurally perfect. The vocals have this amazing haunting emotional grit. The rhythm section is totally locked in. The drop hits exactly where you want it. It is flawless. Until the last 10 seconds. The dreaded fade out. Exactly. The AI singer just suddenly starts mumbling absolute gibberish. And started completely over. But today. But today, you can literally reach right inside that fully rendered audio file and just delete that one single squeal, leaving the rest of the mix completely untouched. It's wild. We're looking at a landscape that's just, you know, shifting under our feet in real time. It really is. The technology is moving so fast that the traditional guardrails are just snapping. So today we are taking a massive deep dive to explore how the rapid advancement of AI music generation is forcing courts, the music industry and individual creators to just rewrite the rulebook from scratch. The total watershed moment. I mean, we are watching this transition from AI being this blunt novelty generator to becoming a surgical, highly integrated production tool. Right. And the friction point is that the legal and administrative frameworks of the music industry, they were built for a reality that just simply doesn't exist anymore. Exactly. So here's the mission for this deep dive. We've got five curated stories today, and we're going to look at this from basically every angle. We have independent benchmark data Tracking the sheer speed of vocal improvements. We've got freshly filed legal documents showing how legacy rights holders are fighting back Yeah, yeah, which is fascinating. Oh, we'll get into that. Yeah Plus we've got industry software updates fundamentally changing the digital plumbing of the business and Brand new desktop tools that alter how you interact with AI audio turning you from a prompter into a literal producer Because every single piece of this puzzle really connects back to that central tension, right? Like how do we manage this incredible capability when the rule book is still literally being drafted? Exactly. If you are making songs in Suno or incorporating AI into your workflows in any capacity, you need to hear this. Let's jump right into the catalyst for all of these shifting rules, which is the tech itself. It is getting incredibly good, incredibly fast. Yeah, and to measure that, we have to look at the ultimate proving ground, which is the blind listening leaderboard. So we're looking at data from artificial analysis and independent benchmarking company and their vocal leaderboard has a brand new champion That's right. The headline here is that ace music v1 Which comes from a company called a studio has actually taken the number one spot on the vocal music leaderboard Wow Yeah, so Suno v6 is now in second place and Suno v6 mini is sitting in third Now just to establish the baseline of what we're actually dealing with here Ace Music is a web app. You access it at acemusic.ai. Our sources don't describe any confirmed Android or iPhone app yet, and there's no public developer API. So it's, you know, it's a very specific consumer-facing web interface right now. Right. It's highly contained at the moment. But what I want to dig into is the methodology, because we see these leaderboards all the time and they can be so subjective. Like, how is artificial analysis actually quantifying better music. Well, what's fascinating is that they've basically digitized the blind taste test at scale. They're relying on a system originally designed for chess, which is the Elo rating system. Oh, right, like matchmaking and video games. Exactly. When you test a large language model on coding or math, you can grade it objectively, right? Like, did the code compile? Was the math correct? But music is entirely subjective. Right. So artificial analysis uses paid validated human listeners. The system gives two different AI music models the exact same prompt. The human listener hears at least 10 seconds of track A and track B without knowing which model produced which, and they just pick the one they prefer. I really appreciate the rooflessness of that ELO system. It's not a grade out of 100, it's purely about who you can beat in a head-to-head matchup. Exactly, it's a zero-sum game. But wait, hold on, let's look at the actual math here. Right now on the vocal board, which covers songs where the model generates both the the singing and the lyrics from a prompt, Ace Music V1 has an elo of 1,160. Suno V6 is at 1,144, but artificial analysis provides a confidence range for each score to account for statistical margins of error. Ace's confidence range runs from 1,139 up to 1,181, and Suno V6's range runs from 1126 up to 1162. Yep, the range is overlap. They overlap heavily. So mathematically, this isn't really a blowout victory at all. It's basically a margin of error coin toss. Why is the industry crowning ace the new king if the math says it could just be a statistical tie? That is exactly the right pushback. From a strictly statistical standpoint, you're right, it is a tie. The benchmarking site itself places both of these models in a shared rank 1 to 3 band. Okay, so it's not just me. No, Ace also has fewer total votes in the system so far, like around 2,100 votes, compared to over 3,000 for Suno v6. We don't have to understand why the industry is reacting this way. In tech, momentum is everything. That makes sense. The fact that a relatively unknown startup, at least in the mainstream generative space, caught up to Suno's flagship model in just a few months is what's causing the panic. It proves that Suno's moat isn't unassailable. So it's the trajectory, not just the snapshot. But it's a trajectory only on the vocal board. Because if we look at the instrumental board, Suno still absolutely dominates. Oh, completely. Vino V6 is sitting at first place there with an elo of 1,148. Ace Music V1 isn't even ranked on the instrumental board. So Ace has essentially specialized entirely entirely in human voice emulation. Which perfectly aligns with their architectural heritage. Right? Ace Studio was founded back in 2018 and they started specifically with singing voice synthesis software. Oh, interesting. Yeah. Their deep neural network research has been focused on the nuances of human vocal chords, breath control, and vibrato for years. So Ace Music, the tool that generates the whole song, is essentially built on top of that highly specialized vocal foundation. Which explains the qualitative differences we're seeing in the reviews. Decrypt did a hands-on review, and they noted that Ace produced much more coherent mainstream pop styles. But the critical detail was how it handled the vocals themselves. They reported way fewer of those weird moments where the singer suddenly mumbles or invents filler words? The hallucinated lyrics problem. Exactly. Why does that happen so much in Suno? Well, in diffusion-based audio models, the AI isn't actually reading the text in a human sense. It's basically predicting the acoustic features that correspond to phonetic tokens. Okay. So, when the model loses the contextual thread of the language, usually during complex rhythmic sections, it just defaults to generating rhythmic vowel sounds. It sounds vaguely musical, but mean effley nothing. Right. The classic sueno gibberish. Exactly. Ace's heritage and dedicated vocal synthesis seems to give it a much tighter grip on phonetic adherents. A reviewer on Quantzlant ran a test backing that up. Now to be fair, this was just one reviewer's test of 28 tracks, so it's not this massive empirical study. But they reported that Ace sang its reference lyrics with 104 out of 104 words correct. Perfect lyric adherence. That is massive for a creator. It is but it's not just the words It's how the song finishes D Cripped pointed out that ace handled song endings Anyone using Suno knows that landing the plane is incredibly difficult Oh, it loves to throw in a bizarre repeating snare or an abrupt tempo change right of the fade-out always But we do have to look at the stylistic biases because these models are fine-tuned differently fundamentally change the output compared to what Suno is doing. It changes everything about the structural integrity of the song. Most direct audio models like Suno use latent diffusion. They're essentially painting the audio waveform pixel by pixel based on patterns they learned during training. Like an image generator but for sound. Exactly. They have a general sense of the shape of a pop song, but they aren't working from a structured blueprint. By generating MIDI first, ACE is establishing absolute symbolic logic. Meaning what exactly? MIDI is binary data. It dictates exactly which note starts when it stops, what pitch it is, and what velocity it has. The AI builds the skeleton of the song in hard data and then renders the acoustic audio on top of that skeleton. Oh wow. Yeah. This is a model to obey structural rules, which is why it doesn't just trail off into chaotic noise at the end of a track. The blueprint tells it exactly when the song is over. That is a brilliant way to solve the hallucination problem. You force the AI to write the sheet music before it's allowed to sing. So for the listener who is perfectly happy with their current workflow, what is the practical application of this arms race? Well, the Synaptic Plunge team actually developed a really smart habit for this. And they have a song where the vocal delivery is the absolute make or break element. They don't just rely on one platform anymore. Right. They run their own blind A versus B test. It is honestly the most rigorous way to produce A.I. music right now. You take your chorus lyrics, you paste them into Suneo, and then you paste them into A.C.E's free tier. You export both generations, rename the files A and B, so you are completely blind, and then you just listen. And you have to check it across different acoustic environments, you know, listen on your studio monitors, listen on your phone speaker, listen in your earbuds. Let your ears pick the take where every single consonant is crystal clear. Just a quick note on the economics of this, QuantSlan notes, ACE has a free tier with 20 credits a day. And paid plans start around $7.90 a month if billed yearly. But you should always read the terms of service because commercial rights for the flagship model require that paid plan. Always check the licensing. But we have to pivot to the wild card and all of this because it fundamentally shifts the legal liability of everything we just discussed. The open source model. Right. A studio along with a partner called Stepfun has released an open source model called A step 1.5. Now this is a separate weaker model, but its code and its weights are entirely open source, meaning anyone can download it and run it locally on their own hardware. Right. says they plan to open source a V1 Flash version later in 2026. This completely changes the power dynamic. When you generate on a web app, the company controls the guardrails. Exactly. The DronX at LData Engineer claim that while Suno actively blocks you from copying a real artist's song or voice, the open source ASTEP 1.5 doesn't stop you at all. We haven't independently audited their local setup, but it brings up the massive elephant in the room. The open source community is essentially saying, we won't build in guardrails, you can do what you want. It's the Wild West. This feels like the Pepsi challenge for AI, but regarding the open source model, just Because a tool doesn't stop you from cloning a voice, that doesn't make it legal, right? Not at all. And this is perhaps the most dangerous misconception circulating in the AI community right now. The lack of a software guardrail is not a legal license. Right. In fact, open source models shift the liability entirely onto you. When you use a corporate web app, you are somewhat shielded by the company's internal compliance. But if you download a local model that has no guardrails and you intentionally prompted to clone a famous famous singer's voice and then you upload that track to a streaming service, you are assuming 100% of the legal liability for violating copyright and the right法 of publicity. Which means the open source revolution could actually be on legally disastrous for individual creators who don't understand the law. Very quickly, yes. But the h, the legacy labels aren't going after the open source developers in their basements yet. they are going after the deep pockets. The fact that this technology is now capable of flawless human vocals leads directly to the legal backlash. So let's look at the amended Elvis lawsuit against Suno. Oh, this lawsuit is a masterclass in how legacy rights holders are modernizing their legal strategies to combat gender devai. Really is. On November 6th, a coalition filed suit against Suno. The plaintiffs include Elvis Presley Enterprises, authentic brands group who manage the Elvis legacy, and Raleigh Music Publishing. And these are not small players. They have endless legal resources. And we need to dissect the two distinct halves of this lawsuit because they attack the AI process from completely different angles. The first half is straightforward copyright infringement. Right. The input phase. Exactly. The plaintiffs allege that Suno systematically scraped their copyrighted recordings from the internet, stored them, copied them, and used them to train their neural networks. They explicitly list 859 works. Classic hits like All Shook Up and Can't Help Falling in Love, plus catalogs from legendary writers like Otis Blackwell and Ben Weissman and Raleigh catalog hits made popular by artists like Chubby Checker. So that copyright claim attacks the input phase of the AI model. But the second half of the suit attacks the output phase and it centers on name, image, and likeness or NIL. The right to publicity. Right. The legal right to control the commercial use of your identity. Historically, protecting a voice was incredibly difficult. Yet common law cases in the late 80s and 90s, like Bette Midler suing Ford Motor Company for hiring a sound-alike singer for a commercial, or Tom Waits successfully suing Frito Lay for the same thing. But those cases were tough to prove, right? Very tough. Because they relied on proving that the public was actively deceived into thinking the realArtist was endorsing a product. And the Elvis estate isn't relying on those old murky common law precedents here at all. They are leaning heavily on a brand new, highly specific weapon, Tennessee's Elvis Act. Yes. This is a 2024 state law written explicitly with generative AI in mind. It modernizes the right of publicity to explicitly protect a person's voice from unauthorized AI imitation, regardless of whether there is an endorsement confusion. It is a landmark piece of state legislation, filling the void left by federal inaction. And the financial stakes are existential for these AI companies. Tell me about it. According to digital music news, the plaintiffs are seeking up to $128.85 million in statutory damages for the copyright claims alone. And then they added another $23.62 million claim for alleged violations of the Digital Millennium Copyright Act. Let's clarify statutory damages because that number is just terrifying. Damages mean that if the plaintiffs win the court can award a set range of money up to $150,000 per infringed work if the infringement is deemed willful even if the copyright owner cannot prove that they lost a single dollar of actual sales. It's designed to be purely punitive. Exactly. And the DMCA claims specifically alleges that during the scraping process, Suno deliberately stripped out the copyright management information, you know, the metadata identifying the owners from the files. Just to show how coordinated this is, a separate rights holder tied to M&M's catalog filed their own suit against Suno right around the same time. It is a total siege. But the breaking news, what elevates this from a standard lawsuit to a highly aggressive legal maneuver is that just one week after the original filing, the Elvis Coalition filed an amended complaint. An amended complaint just one week later. Why do lawyers do that? Like, why not just wait for the trial? You file an amended complaint early in the process to fortify your structure before the defense even has a chance to file a formal motion to dismiss. Ah, boxing the men. Exactly. You are essentially blocking their anticipated escape routes. The plaintiff stated they uncovered a multitude of additional examples of infringement in that single week. And they brought the receipts. They added 42 new examples of what they of what they call infringing output directly tied to to the claims in the case. But then they did something fascinating. They added 76 more outputs tied to Elvis songs that aren't even officially part of the specific claims in the suit. Songs like Blue Suede Shoes and Hound Dog. Right. They explicitly stated they included these extra examples to show the wholesale copyright infringement that Suno engages into this very day. Why flood the court with examples you aren't even suing over? Because they are trying to establish a systemic pattern of behavior. In copyright law, tech platforms often rely on the DMCA safe harbor. The idea that they're just a neutral tool. Right. And if a user does something illegal, the platform is as long as they take it down when notified. But by showing wholesale infringement across dozens of non-claimed tracks, the plaintiffs are arguing that Suno is not a neutral tool. They're arguing that the platform is fundamentally designed, optimized and operated to generate infringing material. Wow. And to really hammer that home, the complaint includes live web links to the specific Suno generations in question. See, this is the part that every creator needs to listen to. Digital music news cited the specific examples that the lawyers flagged. One was a track generated in February 2026. The user titled the track Can't Help Falling in Love, and the generated lyric actively echoed the famous opening line of the original song. Unbelievable. Another example was a track explicitly labeled cover in the description featuring an Elvis style voicing lyrics that closely mirrored the original hit. Now we absolutely must impartially frame this legal reality. Everything we are discussing are unproven allegations in a complaint. Suneo has not lost this case. They have not been found liable. And the source material does not include Suneo's formal response to this amended complaint. We cannot predict a ruling. Of course. This is just an early crucial test case of how laws like the Elvis Act apply to a foundational AI model builder. And I want to stress this heavily too. Nobody is suing the individual users who type those prompts. The target is Suno. But if you are a creator, you need to look at the behavior that drew the lawyer's eyes to those specific tracks. The lawsuit provides a literal map of the red flags. Exactly. The red flags are undeniable using the exact titles of famous hits, generating lyrics that echo or reword original classics, using the word cover to describe an AI generation, and prompting voices that clearly sound like, like, recognizable human beings. It's like suing the printing press manufacturer because a user printed counterfeit money. You can't control every user. But at Playlist Wall on X asks the real question, is the company that builds the generator liable to? And if the court says yes, Yes. That answer dictates how every AI music app on the market functions moving forward. Definitely. But while the courts debate that, creators need a bullet-proof checklist to protect themselves today. The safest approach is absolute originality. On X, a user named AttheoryBezier summarized the prompting strategy perfectly. They said, For client work, don't bet on it. Describe the sound. 70 Soul Rhodes Piano. That is the golden rule. You describe the sonic environment, warm, 90s rockabilly baritone with slapback echo, driving upright bass. You never type the name Elvis. Never. The synaptic plunge team who makes original Suno music constantly operates on a very simple zero tolerance rule. 100% original lyrics written by a human dissent, 100% original titles. And they only ever describe genres and instruments in their prompts. Never real artists. And they never call something a cover unless they've legally secured a mechanical cover license for the underlying composition. The main takeaway from the Elvis suit is that the Legacy Guardians are actively scanning these platforms right now. So describe the sound, keep your words your own, and ensure your workflow is beyond reproach. Which brings us perfectly to our next question. If the courts are too slow and liability is still being figured out by judges, how does the industry actually enforce anything right now? Well, they don't wait for the courts. Exactly. The music industry infrastructure is taking matters into its own hands. We're moving from the courtroom to the digital supply chain. We are talking about baking AI rules directly into the plumbing of the industry itself. This is a massive structural shift in how music is processed globally. To really understand the scale of this, we have to look at the digital supply chain. When you finish a track, you hand it off to a distributor. The provider's job is to package your audio file with its metadata. The song title, the songwriters, the publishing splits, the release date, the territorial rights, and deliver that entire package to the DSPs, the digital service providers like Spotify and Apple Music. And managing that metadata for millions of songs is an absolute logistical nightmare, which brings us to IDOL, a major Paris--based distributor. a proprietary business software platform called Labelcamp to manage all this plumbing. Labelcamp serves idle in more than 40 other business clients globally, including major independent labels like Because Music, Diddow Music, Compact, and PS. We are talking about a backend system that manages over 10 million tracks delivered to 180 different DSPs worldwide. It is a huge chunk of the global independent music catalog. And on November 11th, Labelcamp launched a new feature called the AI connector. The technical foundation of this is critical. It's built on MCP, the model context protocol. Okay, let's unpack MCP because I've heard this acronym floating around the tech space. The source defines it as an open standard that lets AI chat assistants plug directly into outside software and private data. But mechanically, how does MCP actually interface with a massive proprietary labeled database without hallucinating or requiring the AI to be constantly retrained on new financial data. It's super clever. It solves the context window problem. Previously, if you wanted an LLM to analyze your catalog, you either had to fine-tune the model on your data, which is expensive and highly insecure, or paste all your data into the chat box, which hits the token limit immediately. Right, it just breaks. Exactly. MCP acts as a secure real-time bridge. When a label manager asks Claude or chat GPT a question through LabelCamp, the MCP standard translates that natural language query into a secure API call. A query's LabelCamp's internal database retrieves only the specific relevant data points, injects that data into the LLM's context window securely, and the LLM synthesizes the answer. Oh wow. So the AI never actually absorbs the proprietary data into its foundational weights? just reads it temporarily to answer the question. That is brilliant. It means a label can talk to its entire 10 million track catalog in plain English. And this changes everything in two very concrete directions. First, AI is taking over the tedious administrative work. The boring stuff. Right. Instead of a label manager digging through complex, multi-layered dashboards to build a custom spreadsheet, they can just ask a question. Label can't provide some examples like which of my releases spiked in streams this week or do Instagram reels bring more plays than story stickers on this specific single? And because MCP can route multiple data streams, the AI can cross-reference label camp's internal streaming data with external ad campaign results from Meta or TikTok. Or even internal marketing notes stored in a completely different app like Notion. That's incredibly powerful. But the true power is that it goes beyond read-only analysis. It actually has execute privileges. Wait. Excessive privileges? You mean, you can tell the AI fix the songwriter spelling on all five tracks of the CP and move the singles release date up and it just does it? The AI connector will go in, execute all the necessary database updates, and send the new XML or D-DEX metadata feeds out to 180 streaming services. It literally replaces hours of manual data entry with a single conversational sense Idle's operation staff actually tested this day to day during the beta phase and the efficiency gains are massive. Okay, but I have to challenge the privacy aspect. LabelCamp claims this system works with sensitive catalog and financial data, quote, without exposing anything sensitive. But can we really trust a generative AI chat assistant with the labels entire financial and administrative backend? It's a huge question. What happens if the AI hallucinates a command? What if a manager asks it to update a release date and the AI accidentally changes the territorial streaming rights for a massive artist pulling their catalog. Down in Europe. Is the industry moving too fast by giving LLMs execute privileges on supply chain databases? It is a highly valid concern. We are relying entirely on label camps, press claims regarding their data privacy and security architecture here. We have audited their backend. The risk of administrative hallucination is very real when you give an LLM right access to a database. Which is why best practice dictates that the AI should act as a staging tool. It should propose the database changes on screen and a human operator must physically click approve before the data feed is actually sent to Spotify. Okay, so that's the admin side. But the second direction of this update is the one that directly impacts every single creator listening to this. Labelcamp also added a feature that they say helps rights holders retain control over how AI companies use their music. And this is where the legal and the technical collide. Label camp can now send AI transparency tags directly to the streaming services when AI was involved in a release. Interesting. Furthermore, it allows rights holders to set explicit rules within within the metadata and whether their specific recordings may be used for AI training in generation of the future. On X at vocaloor.ai summed up the impact of this instantly. in. Label camp lets labels set AI training rules and send AI tags to DSPs. Now we need to analyze the legal reality of this. These AI training rules embedded in the metadata are essentially a contract and rights tool. It's a way for a label to state in machine readable code whether they grant permission for their recordings to be scraped. Which is exactly the core issue in the Elvis lawsuit. The legacy labels want a mechanism to say, do not scrape. However, it is crucial to clarify the legal limitation here. Our source material does not state how or even if AI companies are legally bound to honor these metadata tags. We cannot claim that a transparency tag is a legally binding enforcement mechanism right now. It is simply the industry attempting to assert control via data structure. Whether a third party AI scraper actually respects that tag or whether a court will punish them for ignoring it is entirely untested. But even if the legal enforceability is up in the air, the structural change is permanent. Independent sueno creator sitting at home using a standard distributor like DistroKid or TuneCore, you need to pay attention because this is a crystal clear preview of where the entire industry is heading. The shift is moving away from a vague checkbox that just asks, did you use AI? To a highly structured binary data field, a transparency tag that travels with your songs file across the entire internet architecture. The digital plumbing is being permanently upgraded to track AI providence. So what is the practical workflow application for our listeners? Well, the Synaptic Plunge team has a fantastic digital hygiene habit to prepare for this exact reality. For every single song they create, they maintain a detailed how it was made receipt. I love this idea. I've started doing it too, and it is a lifesaver. You just keep a simple text file alongside your final wave file. You explicitly write down vocals, generated in sueno v6, lyrics, 100% human written, instrumental arrangement, human choices, mastering, human edited in a DAW. You break down the exact of every element. So smart. Because when your distributor updates their ingestion system to require these granular AI tags, you won't have to guess or remember what you did on a track you rendered six months ago. You have the receipt ready to go. It ensures you can answer the ingestion forms quickly and honestly, guaranteeing that the correct metadata tag goes out to the DSPs the first time. The overarching takeaway here is that distributors are rapidly turning AI-disclosure from an honor system into hard traveling data. Keeping and honest how it was made note for every release is no longer optional. It's a fundamental requirement of modern digital release strategy. Absolutely. Okay, let's take a breath because we just spent a lot of time talking about the massive legal and administrative headaches of generating AI audio. Now we're going to look at a completely different paradigm. We're going to look at how creators are using AI as a software assistant rather than a music generator. Yes, avoiding the copyright pitfalls entirely. Exactly, because the AI isn't making the sound at all. It's just turning the knobs on your synthesizer. We are talking about Anukari 1.1 and the rise of vibe coding. This represents a fascinating evolution in how humans interface with complex music production software. Let's introduce Anukari. This is desktop software, available for Windows, Mac, and Linux. It is not a mobile app. You use it as a plugin inside your digital audio work, or is a standalone application. And it's a software synthesizer. But it's not a normal synthesizer. Usually a synth relies on oscillators, filters, and LFOs generating standard waveforms like SAW or sine waves. Anukhari is entirely physics-based. It's like having a 3D video game engine specifically for sound. It looks so cool on screen. It really does. You build a little virtual world on your screen. You wire together virtual masses and springs, And then you hit them with virtual mallets, or you bow them like a violin string, or you even feed your own external audio into the physics engine to exact the simulation. And the mathematics behind this are incredibly complex. It relies on advanced mass spring damper algorithms to simulate acoustic reality. The simulation calculates the exact nodal points, the tension coefficients of the virtual springs, and how those virtual objects would ring, resonate, and decay in a physical space. And that math becomes the audio. Exactly. And that real-time mathematical calculation becomes the audio that you hear Now you have virtual strings, metal bars, plates, drum membranes and tubes. You know, I have a whole virtual orchestra. Yeah. You can pluck them, strike them, or even blow into them with a new virtual mouthpiece module. They also added something called a modulus sphere, which is this animated sphere that floats through your 3D physics space. Whenever it physically touches one of your virtual masses, it alters the properties of the sound source in real time. We've also added over a thousand new presets, including highly realistic orchestral instruments, build entirely from these physics models, and they've improved the legato and glide functions for playing melodies. And for pricing, it's a free update for existing owners and new licenses are currently half price, so 68 pounds instead of 135 until December 4th. But here's why we are talking about it today. The AI integration in Onakari 1.1 is subtle, but it completely redefines the user interface. This is the vibe coding part. Yes, you can now connect an AI chat assistant like Claude or chat GPT directly to Onakari, but the AI does not generate the audio, it just operates the synthesizer based on your plain language description. Mechanically, how does this work? Is Claude writing C++ code in the background to build a virtual spring? Not quite C++, but it is writing parameter scripts. Evan Mazzeschi, the founder of Onakari, explained this beautifully to Sound on Sound. He said, there's no AI inside Onukari. It just gives your assistant the controls. Okay, so it's a controller. Right. When you type, I want a rusty pipe being hit with a rubber mallet. The LLM translates that acoustic concept into the specific parameter values required by Onukari's physics engine. The AI calculates the necessary mass density, the spring tension, the damping coefficients, the mallet velocity and automatically patches those virtual cables on your screen. This is a fundamental contrast to Suneo. Suneo, the AI is your voice, your drummer, your entire band. It is generating the final audio waveform. With Onukari, the AI is just your hands. It's doing the tedious mathematical patching of cables because you told it to. This ties into a much larger trend that Engadget recently reported on. It's called vibe coding. Vibecoding is a paradigm shift where creators describe the software or the functionality that they want in plain language to an AI chatbot. And the AI writes the actual code to execute it. Vibecoding? It's such a perfect term. Evan Mazzeschi actually vibecoded Onukari itself. He's a former Big Tech software engineer, and he told Music Tech that building Onukari required a thousand line design essay that he fed to an AI, followed by by a thousand iterations where he gave it feedback and direction. A thousand iterations. Yeah, so it's not just pushing a magical make-synth button. It's a massive iterative collaboration between a human's architectural vision and an AI's coding speed. On X, a user named atiCleanAI summarized how disruptive this is to the traditional plug-in market. They said, since nFX for a DAW used to need a dev team or weeks of coding. Now a solo developer with a strong vision can build a complex three-dimensional physics engine by vibe coding it. But I want to push back on this utopian vision. And Gadget issued a severe warning about this exact trend. Because anyone can vibe code a plugin now without actually knowing how to write DSP code, the market is getting flooded. There are mass of fears of AI slop plugins that have a flashy GUI but sound terrible, up all your CPU or are incredibly buggy. The quality control is definitely an issue. And more importantly, there are security risks. In Gadget Sighted, a security study that AI-generated code can often contain deep vulnerabilities. If I download a random, vibed-coded plugin from a forum, could I be compromising my entire studio computer? Absolutely. The vulnerability of AI-generated code in audio plugins is a very real threat. If the AI hallucinates a memory management protocol in C++, that plugin can easily crash your entire session, destroying hours of work. Or worse, it could contain unpatched network vulnerabilities. That's terrifying. It requires the exact same common sense as downloading anything from the early internet. Define the problem you want to solve. Confirm the plugin supports your specific OS. Favor developers who transparently update their software and judge the tool by its stability and sound, not just because it has an AI sticker slept on the interface. So we have this incredible tool where the AI acts as a master sound engineer sitting next to you, you just say, make it sound like a rusty pipe being hit with a rubber mallet, and they patch the cables. How do we filter the good tools from the AI slop? And how do we use this to elevate our tracks? Well, let's give a concrete example, because this is exactly how you make a suno track sound uniquely yours, bypassing all the copyright paranoia. Let's hear it. Suno is incredible pad sweeping under the bridge to build tension. Right. So here is the workflow. You. You export your Suno songs stems. You load those stems into your main audio software on your Mac or Windows computer. Then you add Anukkari onto a brand new empty track. You can browse their thousand presets or using this new update you just type to the the AI assistant. Build a glass bell instrument with a long shimmering decay tuned to the key of C. When the AI does the mathematical heavy lifting of wiring up the physics. Exactly. Then you play a few notes on your physical MIDI keyboard, exactly where you want them, overlaying them onto the Suno drop. And suddenly you have added a bespoke physics-based texture that absolutely no one else on Earth has in their Suno generation. The Synaptic Plunge team relies on this exact philosophy. The AI handles the busy work of patching the parameters and coding the physics, the critical creative decisions, when to play the note, how loud it should be, how it fits the emotional arc of the arrangement, stay entirely human. Let your ears lead. The newest AI in music software doesn't always make the sound. Sometimes it just turns the knobs for you. Use it to explore complex physics engines faster, but keep your ears in charge of the vibe. Which brings us to our final topic today. Anukari showed us how to add unique human flavor to an AI track. But what happens when the AI track you generated is a glaring structural flaw baked right into the middle of it This is the heartbreak of AI music you roll the dice on Suno and you get a generation that is 90% perfect the groove is amazing the tone is exactly what you wanted But in the second chorus the AI singer hits one blatantly wrong sharp note Just one note ruins it right or there's a weird synth bloop right over the emotional climax of the vocal in the past Last, you had to just throw the whole song away and re-roll, hoping you get lucky again You can blur things, you can change the tint, you can cut the photo in half, but you can't move the people inside the picture. The waveform is baked. Exactly. What RIPX does is turn that 2D photograph back into a 3D movie set. It lets you walk behind the singer, tap the guitarist on the shoulder, and tell him to play a C sharp instead of a C. It's like editing reality. It is. RIPX uses AI to completely unbake the audio. First, it takes a finished stereo mix and rips it into separate stems, pulling the vocals, drums, bass, and guitars under their own tracks. But then it goes deeper. It detects the individual notes inside each of those stems. Mechanically, how is it doing this? Because isolating a vocal frequency from a guitar frequency playing the exact same note has historically been nearly impossible due to phase cancellation issues. It relies on highly advanced object-oriented audio processing. The neural network in RIPX analyzes the spectrogram of the audio file. It doesn't just look at the fundamental frequency. It analyzes the entire harmonic series. The overtones. Right. The specific pattern of overtones that makes a guitar sound like a guitar and a voice sound like a voice. The AI maps those harmonic structures over time, disentangles them from the rest of the mix and assigns them to individual visual blobs on the screen. Each note becomes its own individual blob. And you can grab that blob, you can move it up or down in pitch, you can stretch its timing, shorten it or simply hit delete. It is almost exactly like editing MIDI data, but you are manipulating fully rendered real audio. It represents a fundamental shift in our relationship with recorded sound. And Hit and Mix just released RIPX DAW 9, which is a massive update for Windows and Mac desktops. Any technical requirements we should know about? Windows 8 is no longer supported, and if you have a supported in a video graphics card with at least 16 gigs of video memory, The AI processing, the heavy lifting of that harmonic separation is vastly faster. Let's break down what is new in version nine, because the headline changes directly address the most painful parts of editing AI audio. First, the analysis engine is significantly better. It extracts those note blobs far more cleanly from the underlying waveform. This is a crucial improvement because when you manipulate ripped audio, you almost always encounter artifacts. Those watery, warbly side effects. Yeah, those robotic side effects that make the audio sound like it's being played underwater. They occur when the AI fails to perfectly reconstruct the phase of the harmonic series after you move a note. Version 9 drastically reduces those artifacts, preserving the acoustic integrity of the surrounding audio. It also features better bar line detection, so the grid aligns perfectly with the musical measures and improved key and scale detection. That means when you grab a note that is out of tune and drag it up, the software automatically snaps it to the correct musically appropriate note in the scale and it sounds much more natural doing it. They've also added some highly requested effects directly into the interface including a new distortion and a de-esser. The de-esser is an absolute lifesaver for AI producers. A de-esser is an audio tool specifically designed to tame harsh S and sh sounds in a vocal track. Notice sibilance. And AI vocal models are notoriously guilty of this. Oh, they frequently, speedy, harsh, high frequencies. Previously, you had to export the vocal, run it through a separate plugin chain, and bring it back. Now, you can apply the DSR directly onto the specific vocal blob to soften it instantly. They've also improved the Harmony feature, which generates backing vocals from a lead line. It now adds automatic human-like variations in pitch and timing, so the stacked voices don't sound like identical robotic clones. And there's a feature called smaller rips, which is now on by default. It makes the project files five to six times smaller with no loss in overall sound quality, according to the developers, which saves massive amounts of hard drive space. The pricing for version nine is 69 pounds for the standard version, usually 99 pounds. And the pro version is 138, usually 198. We saw a post from a Japanese ex user at Ootari as and we're paraphrasing from Japanese here, but they noted they had owned rip ex pro for over two years and found the update to version nine was totally free for them. So if you are an existing user, definitely check your license portal. Another Japanese retailer on X noted that the core use case for their traditional musician customers is turning two track stereo mixes into stems automatically for practice, for remixing and for learning complex songs by ear. But for the AI creator, rip X is the ultimate repair tool. It really is. So let's walk through exactly how you use this to save your 90% perfect sueno track. Let's look at three highly specific problems and the workflow to fix them. Problem one, the wrong note. Your sueno singer delivers an incredibly emotional performance, but goes slightly sharp on the word love in the second chorus. Here is the step-by-step workflow. First, download the separate stems from your sueno generation, assuming your paid plan offers them. If not, load the full stereo mix into RIPX and let the AI separate the stems for you. Then, Then solo the vocal track. You find that specific blob on the screen that represents the word love. You grab it with your mouse and you gently nudge it down a few cents until it snaps into the correct pitch. And because of the version 9 improvements, a small nudge sounds completely natural and the emotion of the original take is preserved perfectly. Problem 2. The harsh AI sibilance we mentioned earlier. You solo the vocal track, locate the exact blob where the singer makes a harsh piercing S sound, and you apply the new dieter directly to that single blob to soften the high frequencies without dulling the rest of the vocal line. Problem three, a cluttered arrangement moment. Let's say Suno generated a gorgeous intimate acoustic guitar outro, but right on the final chord, the AI hallucinated a weird clashing, high-pitched synth note that ruins the emotional fade-out. You go into the instrumental stem, locate the specific note blob, representing that synth squeak, highlight it, and hit delete. It is gone from the audio file forever, but the acoustic guitar chord underneath it remains perfectly intact, ringing out naturally. It's basically magic, but a critical piece of advice for beginners attempting this kind of surgical editing always fits with the lightest touch possible. Small pitch nudges sound natural, but dragging a note up a full octave will invariably sound like a chipmunk. And when you apply an effect or change a level, You must always compare your before and after at the exact same volume. Yes. The human ear naturally thinks louder means better. It's a psychoacoustic trick. Don't let volume trick you into thinking a harsh EQ move sounds good. Make sure the fix actually sounds cleaner and more musical. Once it does, the fix stems out of RIPX and brings it back into your main mix session. The synaptic plunge team relies heavily on this for their album finishing process. process. , their - On the vocal Famous titles, echoed lyrics, and sound-alike voices, are the red flags that we'll draw Number Three. A. A. I was used is no longer a vague checkbox on a distribution form. It is becoming a structured metadata tag that travels with your release. Keep it honest how it was made note ready for your distributor to ensure clean digital hygiene. Number Four. Some of the most exciting AI in music software doesn't generate sound at all. It just turns the knobs for you. Let AI like Anukhari's assistant help you explore complex physics engines faster, but keep your ears in charge of the final vibe. And number five. When a generator take is 90% perfect, use note-level, object-oriented editors like RIPXDAW to fix the one bad note instead of re-rolling the entire song. Keep the magic, Fix the flaw. That's today's Sculpting Sound. Thanks for listening to Sculpting Sound from the Synaptic Plunge team. If you've got a few more minutes today, check out our other shows, The Daily Plunge and Silicon Side Geeks. See you in the next one.