Scicom-intl/Multilingual-Normalizer

Multilingual TTS text normalizer (written → spoken)

Training pairs for fine-tuning a small LLM as a text-to-speech normalizer: text is a sentence the way people type it (digits, currency symbols, dates, phone numbers, …) and normalized is the exact spoken form, in the same language, with nothing left that a TTS model cannot say.

52,698 rows — 16 monolingual locales and 6 Malaysian code-switched pairs. Every row is digit-free on the spoken side.

from datasets import load_dataset

ds = load_dataset("Scicom-intl/Multilingual-Normalizer")               # text / normalized
sft = load_dataset("Scicom-intl/Multilingual-Normalizer", "sft")       # chat messages, ready to train

{"id": "ms-en-t-000015", "lang": "ms-en", "language": "Malay-English code-switching (Malaysia)",
 "source": "template", "template_id": "7cc461db78", "slots": ["money:ms", "date:ms"],
 "text": "Encik, bil bulan ini RM 66.50 dan due date pada 25-12-2022.",
 "normalized": "Encik, bil bulan ini enam puluh enam ringgit lima puluh sen dan due date pada dua puluh lima Disember dua ribu dua puluh dua.",
 "split": "train"}

Languages

lang	language	number words from	grammar caveat
en	English (Malaysian context, RM/USD)	app.spoken_normalizer	—
ms	Malay	app.spoken_normalizer	—
id	Indonesian	num2words	—
zh	Mandarin (Malaysian context)	app.spoken_normalizer	—
ta	Tamil (Malaysia, RM)	app.spoken_normalizer	—
ta-LK	Tamil (Sri Lanka, Rs/சதம்)	app.spoken_normalizer	—
si	Sinhala	own tables (verbalize.py)	needs native review: -යි and case suffixes; thousands 11–19 and ≥100,000 left to LLM rows
tl	Filipino	own tables	needs native review: linker (-ng/na) applied heuristically; Spanish-derived time/date words only in LLM rows
ar	Arabic (MSA)	num2words + own counted-noun forms	needs native review: gender agreement of bare counts not modelled; dates/times/units only in LLM rows
fr, es, de, it, pt, nl	French, Spanish, German, Italian, Portuguese, Dutch	num2words + locale conventions	dates in the running-text form (am fünfzehnten März, le quinze mars); bare counts avoid 1 (un/une)
pl	Polish	num2words	needs native review: only int/money/percent/decimal/phone/codes deterministic; dates, ordinals, units, years only in LLM rows
ms-en, en-ms, zh-en, zh-ms, ta-en, ta-ms	Malaysian code-switching	app.spoken_normalizer, one language per number	ta-en / ta-ms need native review

Arabic rows use Eastern Arabic digits (٠-٩) in ~30% of the written side.

Code-switching (9,000 rows)

Malaysian speech is not one language per sentence. A sentence carries a matrix language and drops words, phrases and often the number itself into another, and the reading of the digits follows the fragment they sit in, not the sentence:

Encik, bil bulan ini RM250.50 dan due date pada 12/3/2024.
        ↓ Malay clause              ↓ Malay clause
"… dua ratus lima puluh ringgit lima puluh sen … dua belas Mac dua ribu dua puluh empat."

உங்கள் bill RM66, due date 13 March 2027.
        ↓ Tamil clause    ↓ English clause
"உங்கள் bill அறுபத்தாறு ரிங்கிட், due date the thirteenth of March twenty twenty-seven."

That decision is the label. So these rows are not LLM-written: a frame is hand-written with the read-language tagged on every slot ({money:ms}, {date:en}), the value is filled and formatted in that language, and the spoken form comes from the rule verbalizer with the language forced — digit-correct and language-correct by construction. (The normalizer LLM was tried first and is not a usable teacher here: asked for the spoken form of bil anda RM250 it answered "bil anda ringgit malaysia dua ratus lima puluh" — the currency before the amount, which no Malay speaker says.)

Each slot is read together with the carrier words next to it, because the cue is what fixes the reading — 704251 alone is a quantity, nombor rujukan anda 704251 is read digit by digit; 9.50 alone is a decimal, 9.50 மணிக்கு is a time.

pair	matrix + embedded	rows
ms-en	Malay with English (Bahasa rojak)	1,500
en-ms	Malaysian English with Malay	1,500
zh-en	Mandarin with English	1,500
zh-ms	Mandarin with Malay	1,500
ta-en	Tamil with English	1,500
ta-ms	Tamil with Malay	1,500

The two sources, tagged per row

• source: "template" (49,000) — a sentence frame with typed slots ({money}, {date}, {phone}, …) filled with random locale-formatted values; the spoken side is produced deterministically. Monolingual frames are 10 hand-written seeds per locale plus LLM-written ones; code-switched frames are all hand-written. Digit-correct by construction; grammar risk only where the caveat column says so, because slots that are not safe in a locale are never filled deterministically there.
• source: "llm" (3,698) — natural sentences written by an LLM (gemma-4-31b) per category, then normalized by the same LLM with a per-locale prompt and two deterministic few-shot pairs. Kept only if no digit survives, the output is in the locale's script, ≥80% of the non-numeric words are preserved, and the length ratio is sane (the checks field records this). Expect a few percent residual LLM errors. category: "plain" rows are identity pairs (nothing to normalize).

Splits are 90/5/5, by template for template rows (no frame is shared between train and val/test) and by text hash for LLM rows.

Files

• train.jsonl, val.jsonl, test.jsonl — id, lang, language, source, text, normalized, split plus template_id, slots (template rows) or category, checks (LLM rows).
• *_sft.jsonl — the same rows as {"messages": [system, user, assistant]}, ready for chat fine-tuning. The system message is the normalizer prompt for that locale; for a code-switched pair it says the sentence is mixed and that each number is read in the language of the words around it.
• raw/ — the intermediate caches the release was built from: template_pairs.jsonl (all filled frames), templates_llm.jsonl (LLM-written monolingual frames), llm_sentences.jsonl and llm_pairs.jsonl (the LLM rows with their check results, including the ones that were dropped).

Known limits

• Template rows repeat sentence frames; the LLM rows are there for lexical diversity. Do not train on template rows alone.
• The code-switched rows are all template rows: ~30 hand-written frames per pair. They teach the number-reading decision across languages, not open-domain rojak vocabulary.
• Deterministic Sinhala, Filipino, Arabic, Polish and the Tamil code-switched output has not been checked by native speakers.
• Malaysian-context bias throughout: RM amounts, Malaysian phone and IC formats, local service domains (telco, e-wallet, clinic, parcel, ride-hailing).

Provenance

Generated with the synthetic-normalizer pipeline of the Scicom TTS API repo (synthetic_normalizer.{templates_llm,generate,llm_pairs,build}); the deterministic verbalizer for English, Malay, Mandarin and Tamil is that repo's rule normalizer (app.spoken_normalizer). Rebuild or scale:

set -a; source .env; set +a                                   # OPENAI_* for the LLM stages only
uv run --with num2words --with aiohttp python -m synthetic_normalizer.templates_llm --per-locale 60
uv run --with num2words python -m synthetic_normalizer.generate --per-locale 2500 --cs-per-locale 1500
uv run --with num2words --with aiohttp python -m synthetic_normalizer.llm_pairs --per-category 12
uv run --with num2words python -m synthetic_normalizer.build --sft

generate is free (no LLM) and the LLM stages are cached and incremental. Adding a code-switched pair means adding frames to codeswitch.py; adding a locale means extending verbalize.py (number words, currencies, months, units, safe slots) and locales.py (formats, seeds).