Abhisar Mehta

Our current public product is Teen V1, an expressive multilingual speech model. My work centres on finding an economical operating point: balancing serving cost, time to first audio, and the speed of generating the complete utterance.

A faster configuration is not automatically the better one. I care about whether the extra compute buys a useful improvement for the application, and whether a less expensive configuration can still meet its response-time and generation-speed needs.

The public interface supports speech generation, reference-audio voice conditioning, and streaming. The SDK documentation covers synchronous and asynchronous clients, HTTP and WebSocket delivery, and compatibility with existing provider integrations. The team’s launch write-up explains the model and serving approach.

My earlier work at TeenageWorks included Clarifyed, a learning project centred on explanations and visual STEM material. The speech work is a team effort; the public product documentation is the place to follow its current capabilities.

• Neuroplastic Transformer ↗

  Training code for an adaptive transformer, its plasticity controller, and the accompanying paper and figures.

  Public repository
• npviz ↗

  A recorder and dashboard for architectural changes during training, with transformer and ResNet pruning examples. Available on PyPI.

  Public repository
• TTS pipeline ↗

  Scripts for training, exporting and testing a VITS voice, plus browser inference code. Read the training report.

  Public repository

Along the way: YC Startup School India in Bengaluru, and $10,000 each in AWS Activate and Microsoft Azure startup credits. These were cloud credits, not investment funding; Startup School was an event, not participation in the YC accelerator.