Damga — Text Normalization (TN) & Inverse Text Normalization (ITN)

Türkçe README

Damga (damga on PyPI) is an open-source, offline, deterministic text-normalization engine built on weighted finite-state transducers, with Turkish as the default language (languages). It turns what a speech recognizer hears into what people write (ITN), and what a language model or a database writes into what a TTS voice should say (TN):

Why a grammar engine and not a model:

• Deterministic. The same input always gives the same output; every behaviour is a line in a golden file you can read and change.
• Safe on sensitive entities. Phone numbers, IBANs and plates are never padded or "repaired": if the digit count is wrong the words stay words.
• Fast and small. Sub-millisecond per utterance on one CPU core, no GPU, no network, no LLM; the wheel installs with zero dependencies.
• Built for real recognizers. The grammars accept the mixed forms ASR systems actually emit: written digits inside spoken numbers, punctuation in the middle of a phone number, glued separators, mis-heard e-mail domains.
• Two directions, one set of grammars. TN number writing is the exact inverse of the ITN cardinal grammar, so the two directions can never disagree.

Install

pip install damga                 # Linux, macOS, Windows — no other dependency
pip install "damga[build]"        # + pynini, only for writing or changing grammars

The wheel contains the precompiled grammars of all seven languages and a small native runtime (damga._fstrt, OpenFst statically linked), so nothing is compiled on the user's machine and pynini is not needed at run time. Grammar authors install the build extra (pynini has Linux wheels on PyPI; on macOS use conda install -c conda-forge pynini) and run python -m damga.build after editing a grammar. With pynini present, backend="pynini" (or DAMGA_BACKEND=pynini) runs the grammars live; custom category selections automatically use it.

Quick start: your microphone, on your machine

The fastest way to see Damga work is live speech. Mihu Turkish STT runs on your CPU (no GPU, no cloud, no API key), Damga turns what it hears into written text while you speak:

git clone https://github.com/mihu-ai/damga && cd damga
pip install damga turkish-stt sounddevice     # Linux: sudo apt install libportaudio2 first
python examples/mic_streaming.py              # --list shows input devices, --device N picks one

Say "telefon numaram sıfır beş otuz iki yüz yirmi üç kırk beş altmış yedi":

 … telefon numaram 0 5 32 123
[01] STT : telefon numaram sıfır beş otuz iki yüz yirmi üç kırk beş altmış yedi
     ITN : telefon numaram 0532 123 45 67

The line updates while you talk; when you pause the final line is printed. The recognizer model (about 70 MB) downloads once on the first run. On macOS allow microphone access for your terminal (System Settings → Privacy & Security → Microphone). More engines (Mihu STT cloud, ElevenLabs, Whisper, Ollama): examples/.

Usage

from damga import InverseNormalizer, TextNormalizer, Normalizer

itn = InverseNormalizer()
itn.normalize("yüz elli lira")                    # '150 TL'
itn.normalize_batch(["on beş ekim", "yüzde on"])  # ['15 Ekim', '%10']

tn = TextNormalizer()
tn.normalize("Toplam 25.600 adet, %12,5 indirim.")
# 'Toplam yirmi beş bin altı yüz adet, yüzde on iki virgül beş indirim.'

Normalizer(mode="itn").normalize("yirmi beş bin")   # '25.000'

Options (all keyword arguments, all optional):

Writing-style options for the other languages (off by default; plain post-processing, so they work without pynini):

from damga import InverseNormalizer
InverseNormalizer(lang="en", unit_symbols=True).normalize("two hundred kilometers per hour")   # "200 km/h"

Deterministic (default) and non-deterministic TN

By default Damga is deterministic: one written form, one spoken form, every time. That is what a voice agent needs, and normalize() always works this way: about 1 ms per sentence on one CPU core, the same reading in every call. Choosing a reading with a language model would add 0.3–2 s per sentence before the TTS starts, so for voice AI keep the deterministic default.

Some forms are read in more than one way by Turkish speakers (09:30 is dokuz otuz or dokuz buçuk). With deterministic=False, TN also lists those readings, default first. Use it to prepare TTS or ASR training text that matches what a speaker actually said:

tn = TextNormalizer(deterministic=False)
tn.normalize("Saat 09:30'da")             # 'Saat dokuz otuzda'  (unchanged)
tn.normalize_variants("Saat 09:30'da")    # ['Saat dokuz otuzda', 'Saat dokuz buçukta']
tn.normalize_variants("2,50 TL")          # ['iki lira elli kuruş', 'iki Türk lirası elli kuruş', 'iki buçuk lira']
tn.normalize_matching("Saat 09:30'da", heard="saat dokuz buçukta")   # 'Saat dokuz buçukta'

ITN stays deterministic: a phone number or an amount has one written form, and guessing between several is exactly what ITN must not do. Details: docs/deterministic.md.

Streaming ITN for live ASR

from damga.streaming import StreamingNormalizer

sn = StreamingNormalizer(mode="itn")
sn.update("yirmi beş", is_final=False).text              # '25'        (provisional)
sn.update("yirmi beş bin", is_final=False).text          # '25.000'    (still provisional — nothing committed)
sn.update("yirmi beş bin altı yüz", is_final=True).text  # '25.600'    (final, state reset)

update() always receives the latest full hypothesis of the current utterance. A prefix is committed only when it has been stable for stable_after updates and does not end inside a span a grammar could still extend, so a number is never frozen too early. The final result is always the batch normalization of the whole utterance. See docs/streaming.md.

How it works: a token-based, two-pass WFST pipeline

Damga is a token-based normalizer. Normalization is not one big spoken → written rewrite; it is two transducers with a structured, inspectable token string in between:

"toplam yirmi beş bin altı yüz lira kırk kuruş ödedim"
        │
        │  pass 1 — reader FST (union of all readers, weighted)
        ▼
word="toplam" money(integer="25600" fractional="40" currency="TL") word="ödedim"
        │
        │  pass 2 — writer FST (union of all writers)
        ▼
"toplam 25.600,40 TL ödedim"

• Pass 1, reading. Every semantic class (cardinal, money, time, telephone, email, …) is a small grammar that maps a spoken span to a token class(field="value" …). Words no class claims fall through to word="…". The sentence reader is the weighted union of all class readers; the best analysis is the shortest path through the lattice. A semantic class costs about 1, the word fallback costs 100 per token, so any grammar that matches beats leaving the words alone, and among grammars the more specific one wins (iban < email < telephone < money < cardinal).
• Pass 2, writing. Each class has a writer that renders its token: money puts the currency symbol after the amount, time zero-pads and re-attaches the case suffix (saat 09:30'da), telephone groups digits the Turkish way. Fields are emitted in a fixed order by the reader, so there is no Python re-ordering stage between the two passes: the whole pipeline is two compositions and two shortest-path searches.

Why token-based rather than a single "tokenless" transducer that rewrites text directly?

The second composition is cheap (the writer has a few hundred states), so the token layer costs well under a millisecond and buys composability, a readable intermediate form and golden tests at the tag level. The numbers grammar is the foundation: fixed-width, zero-padded sub-graphs (two digits, three digits, thousands, millions …) composed with a leading-zero stripper, reused by every class that contains a number. Ambiguity is resolved with path weights, not code: a non-standard bir yüz carries a penalty, a telephone group ending on a round word (otuz, iki yüz) is penalised so beş yüz otuz iki yüz yirmi üç is read 532 | 123. Details in docs/architecture.md.

Runtime

Grammars are written in Pynini and compiled once into two OpenFst files per language and direction (damga/<lang>/<mode>/compiled/). At run time the native extension damga._fstrt (OpenFst statically linked, ~2 MB) loads them and runs input chain ∘ FST → shortest path in C++; the GIL is released during the search. With pynini installed you can also run the grammars live (backend="pynini"), which is what grammar authors and custom category selections use. Both backends execute the same transducers and are tested to give identical output.

Using Damga with speech recognizers and language models

Damga sits between the recognizer and whatever consumes the text, and between the text generator and the TTS voice:

microphone → ASR (Mihu STT, Whisper, …) → Damga ITN → written text → LLM / CRM / display
LLM (Ollama, …) / database → written text → Damga TN → spoken text → TTS

Mihu Turkish STT

turkish-stt is a real-time, CPU-only Turkish recognizer with turn-end detection. It writes numbers as words; Damga turns them into their written form. Damga does not depend on it:

import mihu_stt                                # pip install turkish-stt (on-prem, CPU)
from damga import InverseNormalizer
from damga.streaming import StreamingNormalizer

itn = InverseNormalizer()
print(itn.normalize(mihu_stt.transcribe("audio.wav")))   # "… sıfır beş otuz iki …" → "… 0532 …"

stt, sn = mihu_stt.StreamingSTT(), StreamingNormalizer(mode="itn")
for pcm16_chunk in audio_frames():             # 16 kHz int16 bytes, as they arrive
    text, is_final = stt.accept_pcm16(pcm16_chunk, 16000)
    if text.strip():
        res = sn.update(text, is_final=is_final)
        show(res.text)                         # best current rendering, tail may still change
        if is_final:                           # recognizer detected end of turn
            handle(res.text)                   # == itn.normalize(text)

The same recognizer is available as a cloud API; its transcripts go through itn.normalize the same way. examples/mihu_stt_onprem.py, examples/mihu_stt_cloud.py and examples/elevenlabs_stt_tts.py (ElevenLabs STT + TTS) are complete scripts; docs/voice-pipeline.md walks through a live call and a batch transcription step by step.

Whisper, Ollama and other pipelines

docs/integrations.md and the examples/ folder show how to put Damga behind Whisper / faster-whisper (batch and segment-wise ITN, and why Whisper's mixed 40 bir output needs it), in front of a TTS voice fed by Ollama or any other LLM (sentence-wise TN on the streamed reply), and inside a generic streaming recognizer loop (Vosk, sherpa-onnx, cloud ASR partials). The core rule for any streaming source: send Damga the recognizer's latest full hypothesis, never a delta.

Using Damga for NLP and data work

Damga is not only for voice products. Anything that reads spoken or transcribed Turkish (or English, German, …) gets numbers, dates, amounts and identifiers in one written form, the form NER models, regular expressions, databases and language models handle best:

from damga import InverseNormalizer, TextNormalizer
itn = InverseNormalizer()

itn.normalize("müşteri yirmi beş bin lira iade istiyor")
# müşteri 25.000 TL iade istiyor
itn.normalize("on dört haziranda saat dokuz buçukta arayın, numaram sıfır beş otuz iki yüz yirmi üç kırk beş altmış yedi")
# 14 Haziran'da saat 09:30'ta arayın, numaram 0532 123 45 67
itn.normalize("plakam otuz dört a b c yüz yirmi üç, e posta adresim ali nokta yılmaz et gmail nokta com")
# plakam 34 ABC 123, e posta adresim ali.yilmaz@gmail.com

• Call-centre and meeting analytics: amounts become numbers you can sum and filter (complaints above 5.000 TL), dates and times become comparable.
• Information extraction: phone numbers, IBANs, plates, e-mail addresses, dates and amounts come out in the shape your NER model or regex expects.
• Search: a call in which on dört haziran was said is found by 14 Haziran.
• Language model input: consistent written numbers, fewer tokens, better arithmetic.
• Fair ASR evaluation: normalize both the reference and the hypothesis, so yirmi beş kişi against 25 kişi is not counted as two errors.
• TTS and ASR training data: normalize_matching writes each sentence the way it was actually said in its recording:

tn = TextNormalizer(deterministic=False)
tn.normalize_matching("Saat 09:30'da 250 TL ödendi.", "saat dokuz buçukta iki yüz elli lira ödendi")
# Saat dokuz buçukta iki yüz elli lira ödendi.

Damga normalizes forms that are written differently from how they are said (numbers, dates, times, money, measures, phone numbers, IBANs, plates, e-mail and web addresses). It is not a general text cleaner: slang, spelling errors and stemming are left to tools such as Zemberek, which work well next to it.

What is covered

ITN: cardinals (with a glued case suffix: yirmi altıda → 26'da), ordinals, negatives, decimals, money (TL/$/€/£ + kuruş/sent), percent/permille, measurements (km, kg, lt, m², °, MB …), clock times (saat dokuz otuz, dokuzu çeyrek geçe, dokuza on kala, case suffixes), dates (on beş ekim iki bin yirmi altı, ekimde, yılında, iki bin yirmi altıda → 2026'da), Turkish phone numbers (any grouping, +90), IBAN, vehicle plates (letters, letter names or the phonetic alphabet), e-mail and web addresses, street addresses (numara on iki daire yedi → No: 12 Daire: 7), the Turkish letter-coding habit (Diyarbakır'ın D'si Edirne'nin E'si … → De…), abbreviations and brands, custom vocabulary. TN covers the same classes in the written → spoken direction, including apostrophe suffixes (2026'da, 25'te, 3'üncüsü).

Known limitations are listed in docs/grammars.md (for example: a lone bir is treated as the article, bare four-digit numbers are not read as years, inflected ordinals such as üçüncüsü are not converted by ITN).

Brands, domains and custom terms

Three data layers, all plain TSV under damga/tr/itn/data/ (see its README for sources and licences):

• brands.tsv — hand-written brand/product names with official casing and Turkish mis-hearings (vatsap → WhatsApp, peypal → PayPal, a ve es → AWS);
• brands_extra.tsv — ~700 well-known businesses, brands and Turkish institutions, common Turkish words filtered out;
• domains.tsv + email_providers.tsv — mis-heard mail domains (cimil → gmail, yahu → yahoo) and ~7 800 free-mail provider labels from free-email-domains (MIT), used to recover a dropped "et" before a known provider. Spellings outside the table (giymiel for gmail) are left as heard: that is a recognizer error, and the fix is contextual biasing (hotwords) in the recognizer with these same lists, not guessing in the normalizer.

Your own terms go in custom_vocab={...} (they win over every table). None of this requires retraining the recognizer; for recognizer-level coverage use contextual biasing (hotwords) with the same lists.

Real recognizer output

Recognizers do not emit clean spoken forms. Mihu STT, for example, produces Sabit hattımız sıfır 5, 3-74, bir dokuz, sekiz yedi bir sekiz or Burak Nokta Koçet örnek banka nokta kom nokta Tier. The ITN grammars therefore accept written digits inside spoken sequences, punctuation inside numbers, glued separators (koçet, noktaçelik, etörnek), mis-heard TLDs (kom, tier, ay) and a table of mis-heard domains (damga/tr/itn/data/domains.tsv).

Measured on 575 real Mihu-STT transcripts of synthetic phone/e-mail prompts (tests/golden/itn_asr.tsv holds 84 of them): telephone numbers 256/256 exact, e-mail addresses 274/319 exact (the rest are recognizer errors such as a dropped "et" or an unrecognisable domain — Damga leaves those as words rather than guessing). On the held-out KPI set (300 phone + 300 e-mail clips, 8 unseen voices, Mihu STT v1.2 with hotwords) exact-match is 84% for phone numbers and 97% for e-mail addresses; the misses are digits dropped by the recognizer, not normalization errors.

End-to-end (TTS → Mihu STT v1.2 → Damga) on 1 200 synthetic sentences spoken by 8 unseen voices: money 100%, percent 100%, dates 100%, times 100%, brands 97%, e-mail 92%, phone numbers 83%, addresses 63%, letter-by-letter spelling 52% — overall 79.0% exact (the previous regex normalizer: 52%, with 0% on money, addresses and spelling). Vehicle plates on a separate 100-clip set: 74%. Of the misses, 86% are recognizer errors (dropped digits, mis-heard letter names); the rest are normalization gaps tracked in tests/golden/itn_asr.tsv.

Safety on ordinary speech is tested the other way round: plain Turkish sentences pass through with no change other than number, ordinal, date, money and unit conversions — the phone, e-mail, plate and spelling grammars never fire on plain text, untouched words keep their casing, and normalize(normalize(x)) == normalize(x). The recognizer frequently emits written numbers ("99 lira", "saat 13.30", "40 bir" = 41); the context grammars accept those mixed forms, bare numbers are left untouched so identifiers are never regrouped.

Tests and benchmarks

pytest                          # unit + golden (TSV) + property + integration tests
python benchmarks/latency.py    # quick per-sentence latency on the shipped runtime (no pynini)
python benchmarks/bench.py --out benchmarks/results/my-machine.md   # full report, needs pynini

Golden sets live in tests/golden/*.tsv (input<TAB>expected); adding a case is a one-line change. Measured numbers are in benchmarks/results/; on a 2026 MacBook running Python under Rosetta (x86_64 conda build, single thread): ITN compile 6.4 s once / cached load 26 ms (5.3 MB FAR), p50 latency 1.1 ms (short utterance), 4.9 ms (two sentences with phone + e-mail), 86 ms (200 words); TN p50 2.2 ms; streaming update p50 1.2 ms; ≈490 MB RSS after a cold compile (less when loaded from the cache). Native arm64 or Linux builds are faster.

Languages

Turkish (tr) is the default language. Grammars live under damga/<lang>/{itn,tn}/; the engine, cache, streaming wrapper and API are language-independent. A new language is a new folder plus one registry line (damga/api.py), after which Normalizer(mode="itn", lang="en"), InverseNormalizer(lang="en") and available_languages() work — see docs/adding-a-language.md.

This release includes Turkish (tr), English (en), German (de), French (fr), Spanish (es), Italian (it) and Russian (ru), TN and ITN for each:

from damga import TextNormalizer, InverseNormalizer
TextNormalizer(lang="en").normalize("EUR 10 million on 6 December 2012")
# ten million euros on the sixth of December two thousand twelve
InverseNormalizer(lang="de").normalize("fünfundzwanzig Prozent")   # 25 Prozent

Readings follow what speakers actually say. The test lines of the languages other than Turkish still need review by native speakers.

Contributing

Grammars are plain Python files under damga/tr/itn/readers, damga/tr/itn/writers, damga/tr/tn/...; data tables are TSV files under .../data/. Failing golden lines are the best bug reports. See CONTRIBUTING.md and docs/architecture.md.

License

Apache License 2.0 — see LICENSE. Pynini and OpenFst are Apache 2.0 as well.