trnorm-js

Turkish text normalization for speech, in JavaScript. Turns written Turkish into the way it is spoken (for text-to-speech), gives a canonical form for scoring speech recognition (WER/CER), and turns spoken numbers back into digits. A port of the Python trnorm with the same output for the same input. No dependencies; runs in browsers and Node 18+.

npm install @ozgurguler/trnorm

import { normalize, normalize_for_eval, inverse_normalize, number_to_words } from '@ozgurguler/trnorm';

normalize("Saat 09:30'da 3. kata gelin, ücret 1.250,50 TL.");
// "Saat dokuz otuzda üçüncü kata gelin, ücret bin iki yüz elli lira elli kuruş."

normalize_for_eval("ABD'de 5 TL mi?");                            // "a be dede beş lira mı"
inverse_normalize('saat on dört otuzda yüzde on beşlik indirim'); // "saat 14.30'da %15'lik indirim"
number_to_words(1923);                                            // "bin dokuz yüz yirmi üç"

What it reads

Numbers, dates, times, money, percentages, units, ordinals, abbreviations, acronyms, phone numbers, IBANs, licence plates, URLs and e-mail addresses, foreign names. Turkish suffixes are re-agreed with the spoken form, which is the hard part a TTS model gets wrong:

Written	Spoken (normalize)
15.00'de	on beşte
%15'lik	yüzde on beşlik
4'üncü, 3. kat	dördüncü, üçüncü kat
1.250,50 TL	bin iki yüz elli lira elli kuruş
5 TL mi?	beş lira mı?
14.10.2026	on dört Ekim iki bin yirmi altı
5 km/sa	saatte beş kilometre
SGK'ye	se ge kaya
Dr. Yılmaz	doktor Yılmaz
iPhone 15	ayfon on beş
0555 555 55 55	sıfır beş yüz elli beş, beş yüz elli beş, elli beş, elli beş
34 ABC 123	otuz dört a be ce yüz yirmi üç

The rules and the reasons behind each choice (in Turkish): docs/TRNORM.md.

API

Function	What it does
normalize(text)	Written → spoken form, for TTS input. Case, punctuation and apostrophes on proper nouns are kept (models use them for prosody).
normalize_for_eval(text)	Canonical form for WER/CER: spoken form, Turkish lowercase, no punctuation, "buçuk" = "virgül beş". Apply it to both the reference and the hypothesis.
inverse_normalize(text)	Spoken → written with digits, for ASR output. Small numbers standing alone stay words ("üç kişi"), as Turkish style guides recommend.
number_to_words(n), ordinal_to_words(n)	4 → "dört", "dördüncü". n is a bigint or an integer; any other number is read as the Python float it is.
tr_lower(text), tr_upper(text)	Turkish-aware case (I ↔ ı, İ ↔ i).

TypeScript types are included. Where the Python raises, this throws (for example a ValueError for a number of more than 4300 digits).

Is it really the same as the Python?

That is the point of it. The port reproduces Python's str, re and int behaviour where JavaScript differs (src/py.js): indexing by code point, Python's Unicode \w \d \s \b, $ also matching before a final newline, unbounded integers (BigInt), Unicode 15.0 case tables (src/case_tables.js), characters newer than Python 3.12's Unicode (src/unassigned.js), and lone surrogates.

• npm test runs 7,566 texts and 2,505 numbers from test/golden.json through every function. tools/golden.py writes that file from the Python trnorm in ahenk-turkish-tts; after a change on the Python side:

  git clone --depth 1 https://github.com/ozguradmin/ahenk-turkish-tts ../ahenk-turkish-tts
  npm run golden   # needs Python 3.12
  npm test

• A wider parity harness lives in ahenk-cep-pdfreader: test suites, real documents and generated hostile input (huge numbers, invisible characters, combining marks, emoji, mixed scripts, lone surrogates, Unicode 16 characters). Its last run matched Python 3.12 on all 742,716 results. The tests pass in V8 (Node, Chrome) and JavaScriptCore (Safari). This library is the in-browser text front-end of that site, Ahenk Okur, which reads PDFs aloud with the Ahenk Cep voice.

Safari 26.0–26.5: their RegExp engine reads a character in U+F800–U+FBFF (or a surrogate) followed by one in U+FC00–U+FFFF (such as fi５) as a single character. For a text with such a pair, the functions throw an error with name === 'EngineError' rather than give a wrong answer; send that text to a server or to the Python instead.

Requirements: a JavaScript engine with ES2022 and Unicode property escapes (Node 18+, current browsers, Safari 16.4+).

Türkçe

Yazılı Türkçeyi okunuş formuna çeviren trnorm'un JavaScript sürümü: tarayıcıda ve Node'da, bağımlılıksız, Python'la aynı girdiye aynı çıktıyı verir. Sayı, tarih, saat, para, yüzde, birim, kısaltma, telefon, IBAN, plaka, URL okunur; ekler okunuşa uyar (15.00'de → "on beşte", SGK'ye → "se ge kaya").

• normalize: yazılı metin → okunuş (TTS girdisi); büyük/küçük harf ve noktalama korunur.
• normalize_for_eval: WER/CER ölçümü için kanonik biçim; referansa ve çıktıya aynen uygulanır.
• inverse_normalize: okunuş → rakamlı yazı (ASR çıktısı için).
• number_to_words, ordinal_to_words, tr_lower, tr_upper.

npm test, Python trnorm'un 7.566 metin ve 2.505 sayıdaki çıktısıyla karşılaştırır; daha geniş eşitlik testinde 742.716 sonucun hepsi Python 3.12'yle aynı çıktı. Kurallar ve tercihler: docs/TRNORM.md. Lisans Apache-2.0. Zorunlu değil ama projenizde kullanırsanız bir bağlantı verirseniz ya da bize haber verirseniz (destek@ozgurguler.tech) seviniriz.

License

Apache-2.0. The Python trnorm, its rules and the Ahenk speech models: ozguradmin/ahenk-turkish-tts.

Not required, but if you use this in a project, we would be glad of a link back or a quick note (destek@ozgurguler.tech).