I got obsessed with exploring voice agents after making the same frustrating phone call over and over again. I thought the hard part would be making the agent intelligent. Instead, I kept getting pulled into all the small systems problems underneath it: latency, transcription mistakes, tool calls, and everything that makes a conversation feel just a little bit worse.
Premove is my open-source effort to work on those problems one by one.
The broader project is still under development. As parts become useful on their own, I’ll release them independently instead of waiting for the whole thing to be finished.
The first released piece is Premove ITN, an open-source, context-aware inverse text normalizer for English voice-agent transcripts.
One of the first problems I hit was that an ASR transcript could look completely fine to me and still be wrong for the tool behind the agent.
my room number is two thirty
→ room_number = 230
meet me at two thirty
→ 02:30The same spoken words can need different written forms depending on context.
That is why I built Premove ITN: to turn spoken values like IDs, phone numbers, dates, times, amounts, and emails into the structured forms downstream tools expect.