Hello, I’m Abhisar.
At TeenageWorks, my focus is cost optimisation: keeping speech generation economical while meeting practical targets for response time and generation speed.
Building with Vyom Patel. Notes on the experiments, measurements, and trade-offs along the way.
Our current public product at TeenageWorks is Teen V1, an expressive multilingual speech model. I focus on the cost of running it alongside the speed and responsiveness people need. More about the work →
The question is not just how quickly audio starts. It is how much serving costs, how quickly the rest of the speech is generated, and which trade-offs make sense for the application.
I started with two layers and a slightly uncomfortable question: should architecture really remain frozen for the entire training run? This is the story of 236 structural decisions, including the useful ones and the questionable ones.
A practical account of data preparation, spot instances, failed runs, ONNX export, and why removing the inference server changed the product.
Why tables and loss curves were insufficient, and how that frustration became npviz.
Teen V1: expressive multilingual speech, with a developer API and SDK documentation.
Adaptive architecture experiment, training code, and technical paper.
Inspect neural networks as their structure changes.
VITS training scripts, ONNX export, browser inference code, and a training report.
I’m Abhisar Mehta, an AI/ML builder interested in the awkward space between an experiment and a real system. I like models that can be inspected, software that can run close to its user, and writing that shows the failed assumptions—not just the final graph.
Find me on GitHub, Hugging Face, or LinkedIn.