Premove AI is an open-source effort to reduce end-to-end latency in conversational voice AI. We build practical infrastructure for real-time voice agents, with a focus on making the path from speech to reliable action faster, more efficient, and easier to run.

Premove ITN is our first public project. More open-source infrastructure for conversational voice AI is in development and coming soon.

Premove ITN is an open-source, context-aware inverse text normalization library for conversational voice-agent transcripts, with open weights. It converts spoken ASR output into written forms for phone numbers, email addresses, identifiers, dates, times, money, measurements, and alphanumeric codes.

the room code is one oh five  →  the room code is 105

Premove ITN separates normalization into three steps: deterministic Rust candidate generation, contextual scoring with one DeBERTa encoding, and exact dynamic-programming decoding.

Install the Python package:

pip install premove-itn

from premove_itn import PremoveITN

itn = PremoveITN.from_pretrained()
print(itn.normalize("the room code is one oh five"))
# the room code is 105

On the frozen 400-row voice-agent benchmark, Premove ITN reached 99.50% semantic accuracy (398/400). The benchmark was held out from training and checkpoint selection. It is a synthetic stress benchmark, not a sample of live production traffic.

• Source code and documentation
• Python package
• Open model weights
• Benchmark report
• Architecture