text2spoken

text2spoken rewrites written text into the words that a person says when they read it aloud. It reads numbers, dates, times, money, measures, percentages, telephone numbers, colours and abbreviations. The rest of the text stays as it is, character for character.

from text2spoken import normalize

normalize("Dr Smith paid £3.50 for 2 tickets on 14 March 2024.", "en-GB")
# 'Doctor Smith paid three pounds and fifty pence for two tickets on the fourteenth of March twenty twenty-four.'

Install

uv pip install --prerelease=allow text2spoken

The library needs Python 3.11 or later. It uses no network and loads no model.

Uses

Text to speech

A voice speaks words, not digits or symbols. Normalise the text before you send it to the voice:

normalize("Meet on 03/04/2025 at 7:30 pm.", "en-US")
# 'Meet on March fourth twenty twenty-five at seven thirty PM.'

The language tag selects the reading. en-US reads 03/04/2025 as month, day, year. en-GB reads it as day, month, year.

Phonemizers

A grapheme-to-phoneme model or a rule set expects words. A digit or a symbol has no pronunciation of its own, so the phonemizer drops it or spells it wrongly. Normalise first, then phonemize:

words = normalize("It is 5 km away.", "en-GB")
# 'It is five kilometres away.'
phonemes = my_phonemizer(words)

Word error rate

An ASR system writes "ten thirty" where a reference transcript writes "10:30". The two texts say the same thing, but a word error rate counts each difference as an error. Normalise the reference and the hypothesis with the same tag and the same params:

ref = normalize("The train leaves at 10:30 on 3 May from platform 2.", "en-GB")
hyp = normalize("the train leaves at ten thirty on the third of may from platform two", "en-GB")
# ref: 'The train leaves at ten thirty on the third of May from platform two.'
# hyp: 'the train leaves at ten thirty on the third of may from platform two'

Text that is already in words does not change. The library does not change case or punctuation, so lowercase both texts and remove punctuation before you score them.

How it reads text

The library finds each written form that it can read, and replaces it with words. It does not guess. A form that it cannot read for sure stays as written, and the library logs a WARNING on the text2spoken logger:

normalize("The run took 1:23:45.", "en-GB")
# 'The run took 1:23:45.'  (a duration, not a clock time)

A second pass changes nothing: normalize(normalize(text, lang), lang) gives the same text as one pass. Output depends only on the text, the tag and the params.

The API

lang is a BCP 47 tag, for example en, en-GB or pt-PT. A tag that is not valid raises InvalidLanguageTagError.

Grammatical agreement

In many languages a number agrees with its noun. analyze and render let the caller give that information. hints maps the start offset of a span to Features, for example a grammatical gender:

from text2spoken import analyze, render

analysis = analyze("Gate 7 opens at 9:15.", "en-US")
[(span.start, span.end, span.form) for span in analysis.spans]
# [(5, 6, <Form.CARDINAL: 'cardinal'>), (16, 20, <Form.TIME: 'time'>)]
render(analysis)
# 'Gate seven opens at nine fifteen.'

Params

NormalizationParams is a frozen dataclass. A field that is None takes the default of the language variety.

A lexicon entry reads a form your way:

from text2spoken import FrozenMap, NormalizationParams

params = NormalizationParams(lexicon=FrozenMap({"en": FrozenMap({"NHS": "N H S"})}))
normalize("Ask the NHS today.", "en-GB", params)
# 'Ask the N H S today.'

To read only some forms, select them:

from text2spoken import Form

normalize("Pay $5 on 12/03/2024 for 3 items.", "en-US", NormalizationParams().only(Form.CARDINAL))
# 'Pay $5 on 12/03/2024 for three items.'

Errors and warnings

Every error is a subclass of Text2SpokenError, and every warning is a subclass of Text2SpokenWarning.

Your own tokenizer

The tokenizer argument takes a callable (text, lang) that returns a sequence of Token(start, end, text). Give it when your pipeline already splits text into tokens, so that the spans and the context words follow your tokens. With no tokenizer, the library uses its default.

Languages

Each language is one file in text2spoken/lang/, named by its primary language subtag. A language with no file reads cardinal numbers from the CLDR spell-out rules, and keeps all other forms as written.

To see what a tag reads:

python -c "from text2spoken import supports; print(supports('en-US'))"

To print the data of a language file as JSON:

python -m text2spoken.export en

English reads the varieties en-GB and en-US. It does not read Roman numerals, letter codes such as GPS, or codes such as X5 after a name; these stay as written. The test rows in tests/data/en.tsv name their source. The checked_by column shows that no native speaker has checked them.

Development

uv venv
uv pip install --prerelease=allow -e ".[test]"
pytest tests

Licence

Apache-2.0. See LICENSE.