The unglamorous plumbing behind Chinese speech recognition

What it does Converts raw Chinese text into speech-friendly forms by rewriting numbers, dates, currency, phone numbers, and fractions into spoken Mandarin—turning “324.75克” into “三百二十四点七五克”, for example. It also strips punctuation and adjusts English word casing to match ASR/TTS lexicon conventions, then emits results in plain text, Kaldi .ark, or TSV formats.

The interesting bit The author is refreshingly blunt about why good text normalization is scarce: it is language-specific, task-specific, and “dirty” work that commercial shops treat as a proprietary asset. Instead of adding another abstraction layer, the repo ships ready-to-run regex-based NSW normalizers and Thrax grammars for inverse text normalization.

Key highlights

• Handles seven Non-Standard-Word types: cardinal, date, digit, fraction, money, percentage, and telephone numbers.
• Reads three input formats natively: plain text, Kaldi archive (.ark), and TSV tables.
• Adapts English casing to align with ASR/TTS lexicon requirements (upper or lower).
• Includes both forward TN (Python) and inverse ITN (Thrax grammar) pipelines.
• Ships as a working module, not a framework that merely “supports” normalization.

Caveats

• The regex-based NSW normalizers suffer from “unintended matches” and need refinement.
• ITN coverage is incomplete; the Thrax grammars handle only a subset of scenarios.
• The repo is explicitly in maintenance mode: the author calls it “done is better than perfect” and does not plan frequent updates.

Verdict Grab this if you are building or maintaining a Chinese ASR pipeline and need a pragmatic normalizer that runs today. Skip it if you want a general NLP toolkit or a framework for other languages; this is tightly scoped to Chinese speech processing.