In the past, I got good results with trying to reduce the variance in entropy in-between tokens, which you can implement very easily by starting with each single character as its own token and then doing a greedy merge of the most numerous outlier token pairs in a loop until you reach your desired token count. https://arxiv.org/abs/2206.12693

I built 3 products from idea to paying customers. So I can help your startup with architecture, implementation plans, and writing specs that you need to keep vibecoding sane. Depending on whether I get into YC (or most likely not, applied with just an idea), I might have free time coming up in the upcoming months.

https://hajo.me/cv/

1. There's plenty of good open TTS models for EU languages: Fish, CosyVoice, Voxtral

2. KugelAudio claims to beat ElevenLabs, but so do Chatterbox-Turbo and Fish Audio S2 Pro.

3. Their 39ms latency is not a strong technical differentiator from Fish at 100ms.

=> It's a deployment business on top of a commodity model. In my opinion, there is no technical moat to defend them from competition.

Their angle is EU sovereignty, which I very much like, but I don't see how their API (where you need to trust them) could be better at sovereignty than an open model which can run air-gapped with no trust needed. But I mean YC is in the VC business so there must be some angle how KugelAudio could 100x their current valuation. Otherwise, they wouldn't be attractive to VC money.

Does anyone know what KugelAudio's unique angle is?

Comments URL: https://news.ycombinator.com/item?id=48358446

Points: 4

# Comments: 1

And I didn’t expect: VR streaming works flawless, too. I just had to buy a WiFi 6 usb dongle.

Remote: Yes (working EU+US remotely since 2014)

Willing to relocate: No, but I travel for kickoffs and on-site work

Technologies:

  AI / ML - PyTorch, TensorFlow, custom C++ inference, RL, speech, computer vision
  Systems - C++ (NEON / AVX), Python, CUDA, kernel drivers (Win / Mac / Linux)
  Distributed - Ansible, Docker, PostgreSQL
  Real-time - DSP, vectorized audio, low-latency pipelines
  Hardware - PCB design (KiCad / Altium), firmware, FPGA, USB3

GitHub: https://github.com/fxtentacle

Email: hn (at) hajo (dot) me

Hi, I'm Hajo. 20+ years shipping production systems, 12+ as CTO. I'm a fractional CTO and I offer AI R&D, performance optimization, and due diligence as freelance consulting services. You might know me from:

ImageRights: sole architect, 10+ years in prod. Web crawling stack (5000+ concurrent tasks via Ansible), distributed C++ content-based image retrieval database scaling to billions of web images, all internal AI filters, plus workflow automation. Still running today.

Spatial Audio Designer: sole developer. 500+ channels real-time on 2013 hardware. NEON / AVX-vectorized DSP, plugins for VST / RTAS / AU / AAX, kernel drivers for virtual sound cards on Mac and Windows, ARM port for embedded and VR. Still in use by famous Hollywood studios.

And yesterday, I figured out how to make MSFS2024 run with multi-camera and instrument pop-outs in Proton 11. Understanding the root cause required tracing the full stack from inside the game, through vkd3d, Proton, the host Linux OS, down to the GPU driver.

Ideally, I'm looking for 5-10 days per month part-time projects. Plus I'd love to work on applied AI / robotics.

npm i -g @musistudio/claude-code-router

So it looks like the person who set up this GitHub repository didn’t even make the software that does all the actual work. They also never mention the original author in the README. But, of course, their AI prompt includes instructions to open their homepage, so this is effectively using the AI prompt for indirect advertisement. The HN submission title is also misleading. I didn’t see any maths in the repo.

EDIT: the underlying tool was discussed here on HN 9 months ago: https://news.ycombinator.com/item?id=44705958

Mediocre programmers use AI to write sed one-liners for them and then brag about AI productivity boosts.

And experts just use sed directly and then they wonder why everyone else is raving about AI.

because companies will need

“proof of productivity gains or metrics that show a clear return for all this AI investment.”

which in my opinion is simply not true. I haven’t seen any good study that showed AI to actually improve productivity overall. It massively helps in some areas, but then promptly gets stuck in others. So you still need an expert to guide it.

Also, it appears that >99% of feminism researchers are publishing their scientific papers with a feminine name. I can easily understand why the general public might confuse the 2 groups with each other.

Which brings me back to the question: what do you think the authors hope to gain by invoking this association? Especially now that we have established that their word choice is highly likely to be misunderstood?