I'm building a Klatt-style formant speech synthesizer targeting behavioral compatibility with a specific commercial synthesizer. The architecture is driven by hundreds of academic papers, uses declarative YAML with CEL expressions for phased rule execution, and every parameter is citation-backed to specific phonetics research[0].
Nobody is selling this off a shelf. The market for "Eloquence-compatible formant synthesizer with a citation-backed parameter space" is approximately me.
I could not have built this without AI coding tools, and I've been a professional developer for 15+ years. I wrote the specs, I chose the architecture, I read the papers, I know what the output needs to sound like. The sheer surface area of translating hundreds of papers worth of acoustic phonetics into a working runtime would have taken me years solo. With Claude Code it's taken months, and I'm still the one catching when it misinterprets a Klatt coefficient or botches a formant transition rule, because you have to actually know the domain to do that.
Reducing what you do not understand to "Ikea or microwave meals?" Because you don't like it? or aren't familiar with it? Is saying a thing about you. Not about people who know how to use the tools.
I think you're missing my point. It has nothing to do with what I like.
I'm saying there's a ton of nuance and human feedback necessary to build the software most developers work on for a living. It's built to requirements that evolve with the business. Businesses ultimately serve people with opinions and preferences. Businesses need to pass audits. Specs can change quarterly. Clients and their contracts come and go. It's exactly like building/maintaining custom furniture for a bespoke house, or consistently cooking a signature recipe at any scale and considering any necessary accommodations. If it wasn't true, the business wouldn't be viable regardless if AI is used or not. I'm talking about systems and services, not products.
You are building a product for a narrow use case requiring DSP. Modeling was always the point. I don't doubt AI helped, but we're not talking about the same thing.
If you were building something for an enterprise client, nobody would give a shit how you got it done as long as you have a demo by monday and it ships next month. No excuses and no gotchas. Any incident risks breaking the SLA and losing the contract. If the client calls a meeting at the last minute to make changes, you're probably working on the weekend. People don't like using AI for stuff like that. Efficiency is not the priority. People don't want cutting edge or novel. They want reliability and competence. Their livelihood depends on knowing exactly how it works and how it can be extended and maintained. They want to stay at least one step ahead of what the client asks for next. People want to test the hell out of it and nobody wants to get fired for not noticing what is obvious to other stakeholders who don't share their tunnel vision. Clients don't have much tolerance for delays or bugs. This is why AI has mixed or negative results for all but personal projects or startups.
Nobody is selling this off a shelf. The market for "Eloquence-compatible formant synthesizer with a citation-backed parameter space" is approximately me.
I could not have built this without AI coding tools, and I've been a professional developer for 15+ years. I wrote the specs, I chose the architecture, I read the papers, I know what the output needs to sound like. The sheer surface area of translating hundreds of papers worth of acoustic phonetics into a working runtime would have taken me years solo. With Claude Code it's taken months, and I'm still the one catching when it misinterprets a Klatt coefficient or botches a formant transition rule, because you have to actually know the domain to do that.
Reducing what you do not understand to "Ikea or microwave meals?" Because you don't like it? or aren't familiar with it? Is saying a thing about you. Not about people who know how to use the tools.
[0]: https://github.com/ctoth/Qlatt/