Skip to main content
Premove ITN separates inverse text normalization into three steps:
The v0.3.0 public API adds a deterministic fourth boundary after decoding:
The first three stages produce the readable normalized text. Enrichment uses only explicit caller context and cannot add candidates, change a decoder edit, or trigger another model pass.

​
Generate

Deterministic Rust realizers enumerate structurally valid written forms for spans in the transcript. The model cannot invent an output that the candidate layer did not generate.

​
Score

DeBERTa encodes the complete sentence once. It gives each candidate a score based on the source span, proposed replacement, candidate kinds, and sentence context.
Both replacements are generated for this span. The frozen v0.2.0 model prediction for the room code is one oh five is the room code is 105. This does not imply that every ambiguity is resolved correctly; see the benchmark limitations.

​
Decode

Candidates can overlap. Exact dynamic programming selects the highest-scoring compatible set of edits across the complete sentence. Leaving source text unchanged is always available through a zero-score KEEP path. The rules decide what can be written. The model decides what fits the context. The decoder decides which edits can coexist. The temporal resolver then adds a canonical value only when the selected span and caller context make it deterministic.

​
Enrich temporal values

normalize_structured() retains the decoder’s exact source and normalized offsets, then applies deterministic contextual rules to compatible DATE spans and unchanged temporal text. For example:
The readable text remains call me tomorrow. The structured span carries resolved_value="2026-09-21". Without a reference datetime, the span remains annotated but unresolved. normalize(), normalize_resolved(), and normalize_structured() share the same internal result and model inference.

​
What this design does not do

The scorer cannot generate arbitrary new strings. An unsupported format needs a Rust candidate before the model can select it. The decoder can also keep the source text when no candidate wins. This avoids forcing a rewrite for every number-like phrase, but it means a supported phrase is not guaranteed to change in every context. Read the implementation architecture for tensor, alignment, scoring, and decoding details. See supported forms for the current public boundary.