Hey everyone 👋 I started building Premove ITN while working on a voice agent. The problem looked simple at first: turn spoken forms like “one hundred twenty three” into 123. But the harder cases were contextual: meet me at two thirty → 2:30 room number two thirty → 230 I first tried deterministic rules, then NVIDIA Thutmose, then a BIO-tagging approach. What finally worked was separating the problem into three parts: 1. Deterministic code generates structurally valid candidates 2. A contextual neural model scores them 3. Exact decoding chooses a compatible final path That became Premove ITN. On a frozen 1,500-row synthetic benchmark, Premove reached 89.7% overall semantic accuracy, compared with 59.4% for NVIDIA Thutmose and 55.8% for text-processing-rs. On the dedicated voice-agent subset, Premove reached 99.5%. The code and model weights are open source. I’d especially love feedback from people building voice agents: what normalization failures are you still seeing in production?

Premove ITN

Open-source context-aware ITN for voice agents

Open-source context-aware ITN for voice agents