spokenform

spokenform converts plain written text in one selected language into reviewable text intended for speech systems. It is a text-to-text frontend: written text in, spoken form out.

The package provides:

• context-aware abbreviation and source-aligned numeric-unit expansion through abbr2words;
• a locale-aware structured-value stage for quantities, dates, times, currencies, temperatures, labels, and contextual ordinals;
• locale-policy-wrapped number, date, time, currency, decimal, and ordinal verbalization;
• optional provider-neutral spaCy annotations for POS-aware abbreviation rules;
• stage-level provenance through PreparedText;
• composed input-to-output offset maps with left/right boundary bias;
• caller-defined and automatically discovered protected ranges;
• conservative protection for URLs, email addresses, and arbitrary multi-dot semantic-version/ID sequences.

High-confidence URL, e-mail, semantic-version, and contextual Roman rendering is available explicitly with normalize_literals=True; caller-protected spans remain absolute. Structured fractions, identifiers, operator-shaped math, music-context tokens, and controlled biological names use semantic precedence and source-aligned mappings.

It intentionally does not detect languages, parse or render SSMD, segment mixed languages, generate phonemes, or depend on kokorog2p.

Installation

python -m pip install spokenform

For optional spaCy integration:

python -m pip install "spokenform[spacy]"
python -m spacy download de_core_news_sm

spaCy and its trained pipelines are separate packages. spokenform never downloads a model automatically.

Quickstart

from spokenform import prepare

prepared = prepare(
    "Prof. Klein bringt am 14.05.2026 um 18:20 Uhr 2 kg mit.",
    language="de",
)

print(prepared.spoken_text)
print(prepared.render_changes())

The result contains:

• source_text: unchanged caller input;
• clean_text: plain text used by the normalization pipeline;
• spoken_text: readable normalized output;
• language: normalized processing-language code;
• ordered stages and mapped edits;
• a composed offset_map;
• structured warnings.

PreparedText.text is an alias for spoken_text.

kokorog2p adapter

Use prepare_for_kokorog2p(text, language=...) for one explicitly selected language run. The adapter preserves caller-owned run whitespace and protected overrides, emits exact source-coordinate replacements, and leaves tokenization, G2P, phonemization, and model punctuation to kokorog2p. German, French, Spanish, Italian, Portuguese, Czech, and English are parity-gated semantic migration targets. English is active on the kokorog2p spokenform adapter for reviewed structured semantics, contextual single-dot release labels, and safe ordinary-number categories; phoneme-sensitive years, suffix ordinals, Roman numerals, phone/ID and arbitrary multi-dot sequences, numeric suffixes, and G2P decisions remain downstream in kokorog2p.

German quantity symbols are recognized by abbr2words.iter_unit_matches(). spokenform owns only the canonical German grammar that realizes those matches, including gender, invariant Stück, currency decomposition, and lexical decimal digits. French likewise realizes canonical abbr2words quantity and currency identities, including French dates, times, ordinals, decimal digits, plural grammar, temperatures, and major/minor currency units. Spanish realizes canonical quantities, temperatures, currencies, dates, and ordinary numbers; Spanish 18:20-style time expressions remain caller-managed. Italian realizes reviewed dates, quantities, temperatures, currencies, and ordinary numbers; Italian colon times remain caller-managed. Portuguese realizes reviewed dates, quantities, temperatures, currencies, and ordinary numbers; Portuguese colon times remain caller-managed. Czech realizes reviewed dates, ordinary numbers, quantities, temperatures, currencies, and canonical extended units; Czech colon times remain caller-managed. No locale copies raw symbol inventories or downstream tokenizer/phoneme rules.

Language boundary

Each call processes one language. Production callers should always pass language=...; English remains the API default for compatibility and simple CLI usage.

Language detection and mixed-language handling belong in the orchestration or G2P layer. A foreign word may remain unchanged through normalization and be handled afterward. Existing source spans can be transferred with prepared.offset_map.

Markup must also be parsed outside this package. Pass plain text to prepare() and use ProtectedSpan for ranges generic normalization must not change.

Configuration

from spokenform import PreparationConfig, prepare

config = PreparationConfig(
    language="en",
    expand_abbreviations=True,
    expand_structured=True,
    expand_numbers=True,
    normalize_whitespace=True,
    context=True,
)

prepared = prepare("The board is 2 in. wide.", config=config)

When a PreparationConfig is supplied, it is authoritative for pipeline options.

Residual symbols and acronym case

Residual punctuation and symbols are unchanged by default:

PreparationConfig(language="en", symbol_mode="none")

Use symbol_mode="remove" to remove all residual Unicode punctuation and symbols, or use an exact-codepoint allowlist with symbol_mode="keep":

PreparationConfig(language="en", symbol_mode="remove")
PreparationConfig(language="en", symbol_mode="keep", keep_symbols=":;,()-,.")

The filter runs after semantic recognition and does not modify protected spans. For generic uppercase acronyms, generic_acronym_case="lower" renders ABC as a b c; the default and "upper" render it as A B C. Lexical acronyms, preserved terms, and known initialisms retain their existing policies. The API policy reference explains the generic_acronym_mode, registered_acronym_mode, and long_number_mode choices, including their false-positive tradeoffs.

spaCy support

spaCy supplies POS annotations for abbreviation rules that opt into POS guards. The public normalization API remains provider-neutral.

abbr2words accepts POS annotations, but its bundled registries do not necessarily require POS labels. Therefore installing spaCy alone may not change default normalization output. The integration is usable for custom POS-guarded entries.

Load and inject a pipeline in the application:

import spacy
from spokenform import prepare

nlp = spacy.load("en_core_web_sm")
prepared = prepare(
    "The board is 2 in. wide.",
    language="en",
    nlp=nlp,
)

Or ask spokenform to load an already installed model:

prepared = prepare(
    "The board is 2 in. wide.",
    language="en",
    spacy_model="en_core_web_sm",
    strict=True,
)

Model names and paths are passed to spacy.load(). Loaded models are cached by language/model key. reset_spacy_cache() clears that cache.

The adapter reads the token attributes text, idx, pos_, tag_, lemma_, and lang_. lang_ is carried as provider metadata; it is not used as language detection. Annotation spans are validated against the exact input text and remapped when protected ranges are replaced by internal sentinels. A trained pipeline with POS or morphological annotations is required for quality improvement; spacy.blank(...) supplies tokenization but normally no useful POS tags.

Explicit annotations take precedence over nlp and spacy_model.

Protection

Use ProtectedSpan(start, end) or a (start, end) tuple to protect a source range:

from spokenform import ProtectedSpan, prepare

text = "Keep Dr. literal, but verbalize 12."
start = text.index("Dr.")
prepared = prepare(
    text,
    language="en",
    protected_spans=[ProtectedSpan(start, start + 3)],
)

Invalid or overlapping ranges warn by default and raise ProtectionError with strict=True. URLs, email addresses, and semantic versions are protected automatically.

Offset mapping

from spokenform import prepare

source = "Prof. Klein has 2 kg."
prepared = prepare(source, language="de")

start = source.index("Prof.")
end = start + len("Prof.")
spoken_start, spoken_end = prepared.offset_map.map_source_span(start, end)

print(prepared.spoken_text[spoken_start:spoken_end])

Use bias="left" or bias="right" when mapping an individual boundary at an expansion.

CLI

spokenform --lang de "Prof. Klein hat 2 kg."
spokenform --lang de --changes "Prof. Klein hat 2 kg."
spokenform --lang de --json "Prof. Klein hat 2 kg."
spokenform --lang en --spacy-model en_core_web_sm --strict "The board is 2 in. wide."
echo "The value is 2." | spokenform --lang en

Examples

Executable examples are in examples/:

python examples/basic.py
python examples/german.py
python examples/german.py --spacy-model de_core_news_sm
python examples/protected_text.py
python examples/offset_mapping.py

Interactive notebook

Try spokenform in your browser without installing it locally:

Launch the spokenform playground on Binder

The Binder notebook runs the selected repository revision and includes interactive controls for language, semantic stages, residual-symbol handling, and generic acronym casing. Changes made in a Binder session are temporary.

Documentation

Documentation sources use MyST Markdown. No reStructuredText source files are required.

python -m pip install -e .
python -m pip install -r docs/requirements.txt
sphinx-build -W -b html docs docs/_build/html

Development

python -m venv .venv
. .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -e ".[dev]"

python -m pytest
python -m ruff check .
python -m ruff format --check .
python -m mypy spokenform examples
python -m build
python -m twine check dist/*

On Windows, activate the environment with .venv\Scripts\activate.

Current limits

• One processing language is supported per call.
• Language detection, mixed-language segmentation, and language marking are external.
• SSMD and other markup must be parsed before calling spokenform.
• Date, time, currency, and ordinal grammar is conservative and not exhaustive.
• Spanish parity ownership covers reviewed dates, quantities, temperatures, currencies, and ordinary numbers; time expressions remain caller-managed.
• Italian parity ownership covers reviewed dates, quantities, temperatures, currencies, and ordinary numbers; colon times remain caller-managed.
• Czech and English own reviewed structured and safe plain-number categories; Czech colon-time candidates remain caller-managed. English owns contextual single-dot release labels such as bot 2.0 as bot two point oh, while ordinary decimals retain digit-wise zero wording. English years, suffix ordinals, Roman numerals, phone/ID and arbitrary multi-dot sequences, numeric suffixes, and G2P decisions remain downstream in kokorog2p.
• abbr2words currently exposes final expanded text rather than semantic replacement objects, so stage edits are reconstructed deterministically from diffs.
• Trained spaCy pipelines must be installed and version-compatible with the spaCy runtime.

Dependency direction

abbr2words ─┐
            ├─ spokenform ── kokorog2p
num2words ──┘

spaCy is an optional quality dependency. spokenform remains independent of language detection, markup parsing, and phoneme generation.

Release versioning

setuptools-scm derives versions from Git tags and writes spokenform/_version.py during builds. Create an annotated tag for the target release, for example vX.Y.Z. The source-tree fallback when SCM metadata has not been generated is the neutral version 0+unknown; release builds derive their version from the annotated tag.

Before publishing, ensure the released abbr2words>=0.2.9,<0.3.0 prerequisite exists on the target package index and run the checklist in docs/release-checklist.md.

License

Apache License 2.0.

Recognition modes and domains

The runtime interpretation policy is separate from rendering options:

from spokenform import prepare

result = prepare(
    "The final was 3-2 and the sample contains H2O.",
    interpretation_mode="surface",
    disabled_domains={"chemistry"},
)

interpretation_mode="contextual" is the default and preserves the existing contextual behavior. surface is fail-closed: only recognizers with intrinsic evidence may claim a structured expression, so ambiguous context-dependent forms can remain unchanged. disabled_domains independently suppresses semantic families such as chemistry, biology, sports, or finance. Use allowed_domains for a fail-closed permitlist that remains stable when future domains are added. sequence_fallback_mode="preserve" is the default; "spell" provides conservative orthographic coverage for residual sequence-shaped spans without spelling ordinary prose. The legacy context option controls abbreviation context and is not the global interpretation mode.

Optional Lexhint evidence

Lexhint can be supplied explicitly when lexical or positive semantic evidence is available:

python -m pip install "spokenform[lexhint]"
lexhint dataset download en --variant runtime

The optional extra supports Lexhint 0.1.2 <= x < 0.3.0, including the Lexhint 0.2.x family. Lexhint artifacts are versioned independently of Spokenform: Lexhint 0.1.x uses schema 7, while Lexhint 0.2.x requires schema-8 artifacts.

For Lexhint 0.2.x, install a current schema-8 runtime dataset explicitly with lexhint dataset download <language> --variant runtime. Spokenform never downloads Lexhint data automatically; it only uses an installed artifact supplied by the caller.

The runtime provider boundary stays narrow and deterministic. Spokenform uses exact lexical evidence, segmentation, and positive semantic-domain corroboration; Lexhint fuzzy and dictionary-search APIs are not used for automatic recognition.

from lexhint import Lexicon
from spokenform import prepare

lexicon = Lexicon("en", variant="runtime")
result = prepare(
    "Visit chatgpt.com.",
    language="en",
    normalize_literals=True,
    lexical_evidence=lexicon,
    use_spacy=False,
)
print(result.spoken_text)
# Visit chat g p t dot com.

The provider language must match Spokenform's base language, so en_US and en are compatible but de is rejected. Lexical-only providers can improve URL rendering; semantic evidence is optional and unavailable semantic capability is not negative evidence. Semantic evidence is used only in contextual mode. Surface mode ignores it, while lexical evidence for an already-recognized URL is still usable for rendering.

Lexhint remains below the interpretation layer. Spokenform owns structured candidate recognition, precedence, domain policy, URL syntax, and speech rendering. abbr2words remains the owner of ordinary prose abbreviation and initialism expansion. Lexhint is not a generic prose acronym detector.