urdu-realtime-1: a realtime speech model for conversational Urdu
neu is training urdu-realtime-1, a speech-to-speech model for natural conversation in Urdu. The model replies in less than 200 ms.

Neu research and engineering · 5 published notes
First-party engineering notes from training our own voice models, speeding up local inference, fine-tuning small tool-calling models, and testing new interfaces.
neu is training urdu-realtime-1, a speech-to-speech model for natural conversation in Urdu. The model replies in less than 200 ms.

Magic Tap combines gaze estimates and tap sounds to select six large buttons. The experiment uses the camera and microphone already in a laptop.
We compared costs for one support agent that processes 800 tickets per month. The comparison includes platform fees, model fees, and self-hosting costs.
We added a prediction head to Ornith 35B. Code generation increased from about 56 to 77.9 tokens per second on a Tesla P40.
We trained two small models for ScallopBot. One improved tool selection. The other reduced memory processing time, with lower teacher agreement.
Talk with the team that trains, benchmarks, and ships these systems.