sherpa-onnx-en-itn

English ITN grammar for on-device SR

Builds en_itn.fst, an inverse-text-normalization rewrite consumed by sherpa-onnx's rule_fsts/rule_fars ITN stage. It turns spoken-form numbers in the recognizer output into written form ("two thousand twenty six" → "2026", "five dollars" → "$5", "ten percent" → "10%").

This exists because our on-device NeMo/Nemotron model spells numbers out, and neither WeTextProcessing (Chinese-only ITN) nor NVIDIA's NeMo-text-processing FARs are usable here — NeMo's exports are a tagger+verbalizer pair that require Sparrowhawk's two-pass parser, whereas sherpa applies each FST as a single-pass rewrite. So we author our own.

The sherpa-onnx FST contract

sherpa loads the FST and runs compose(input_text, rule) → shortestpath. The rule FST must therefore be:

• standard arc type (StdArc / tropical) — asserted on load, log rejected
• vector or const OpenFST binary
• byte alphabet, no symbol table — built with pynini's byte lib
• total over byte input — the grammar is applied as a weighted whole-sentence tagger (space-delimited tokens, each a number-phrase or passthrough word), so unmatched spans stay identity; a non-total FST returns "" and wipes the transcript

build_en_itn.py satisfies all four. test_en_itn.py validates by replicating sherpa's compose+shortestpath, so a green test means it works at runtime.

Build

pynini cannot build natively on macOS arm64, so the build is containerized:

./build.sh

Outputs tools/itn/out/en_itn.fst (and en_itn.far), runs the test suite.

Hosting

Distributed exactly like the punctuation model — a Hugging Face repo under our org, pulled by the on-device ModelManager via OnDeviceModelSpec::Repo. The SDK references it through the ItnModel config (OnDeviceAsrConfig::with_itn) and sets config.rule_fsts in sdk/multimodal-asr-on-device-client/src/engines/asr.rs.

Upload the built FST to the repo the SDK is wired to (brady-pplx/sherpa-onnx-en-itn):

pip install -U huggingface_hub
huggingface-cli login                      # needs write access to brady-pplx
huggingface-cli upload brady-pplx/sherpa-onnx-en-itn tools/itn/out/en_itn.fst en_itn.fst

ItnModel.revision resolves as a branch/tag ref (defaults to main), not a raw commit hash, so to freeze a version create a tag on the repo and set that. The grammar is authored here from scratch, so the artifact carries no upstream license obligation.

Coverage / limitations

Covered: cardinals 0–999,999,999, US-dollar amounts, percentages.

Follow-ups: year readings ("twenty twenty six" currently → "20 20 6"), cents ("five dollars fifty"), ordinals, dates, phone numbers.