ovos-lang-parser

Map spoken and written language names to standard IETF/BCP-47 language codes, and back, in many languages, offline, with a two-function API.

"Brazilian Portuguese"  ->  "pt-br"
"alemão" (Portuguese)   ->  "de"
"pt-br"  ->  "Português do Brasil"   (rendered in Portuguese)
"pt-br"  ->  "Brazilian Portuguese"  (rendered in English)

The library understands language names written in 21 languages (see Coverage). A user can say "French" in English, "français" in French, or "Französisch" in German, and each resolves to the code fr. This is the piece you need whenever a human names a language in free text and your code needs a canonical code to act on: routing to a translation/TTS/STT engine, tagging an entity, or normalizing a messy label.

It ships as part of OpenVoiceOS, but has no OVOS runtime dependency and is useful in any plain-Python project.

Install

pip install ovos-lang-parser
# or
uv add ovos-lang-parser

Runtime dependencies are small: langcodes for tag normalization and ovos-utils for the fuzzy matcher. There are no models and no network calls: the wordlists are bundled.

30-second quickstart

from ovos_lang_parser import extract_langcode, pronounce_lang

# name -> code (second arg is the language the text is written in)
print(extract_langcode("Brazilian Portuguese", "en"))          # -> ('pt-br', 1.0)
print(extract_langcode("translate this to German", "en"))      # -> ('de', 1.0)

# code -> name, rendered in a chosen language
print(pronounce_lang("de", "en"))   # -> German
print(pronounce_lang("de", "pt"))   # -> Alemão

That is the whole surface for most callers: extract_langcode reads a name out of text and returns (code, confidence). pronounce_lang turns a code back into a human name.

The API in one screen

In both functions lang is the language the names are written in, not the language being named. extract_langcode("français", "fr") and extract_langcode("French", "en") both give fr. See docs/api.md for full signatures, the confidence model, and edge behavior.

Use it outside OVOS

The same two functions cover a range of standalone jobs. Each example below is a runnable script under examples/: run pip install ovos-lang-parser, then run the script. No OVOS stack is required.

Entity extraction / NER: pull a language out of free text

You have a sentence and want to know which language it mentions.

from ovos_lang_parser import extract_langcode

for text in ["translate this to Brazilian Portuguese",
             "can you say it in Mandarin Chinese?",
             "I'd like the subtitles in Greek"]:
    code, conf = extract_langcode(text, "en")
    if conf >= 0.7:
        print(f"{text!r} -> {code} ({conf:.2f})")

Because matching is fuzzy, apply a confidence threshold to decide whether a language was really mentioned. Full script: examples/ner_language_mentions.py.

Routing: pick a translation / TTS / STT engine by name

A user names a target language. You resolve it to a code and hand that to whatever engine your pipeline drives.

from ovos_lang_parser import extract_langcode

def resolve_target(user_request, spoken_in="en"):
    code, conf = extract_langcode(user_request, spoken_in)
    return code if conf >= 0.7 else None

print(resolve_target("read it back to me in German"))   # -> de  (feed to your TTS)

Full script (with a mock engine table): examples/routing.py.

Normalization: canonicalize messy names and autonyms to one code

Aliases, autonyms, and localized spellings all collapse to a single canonical code, so you can deduplicate and standardize labels regardless of how they were written.

from ovos_lang_parser import extract_langcode

labels = [("Deutsch", "de"), ("alemão", "pt"), ("German", "en"), ("allemand", "fr")]
for name, written_in in labels:
    code, _ = extract_langcode(name, written_in)
    print(f"{name:>10} -> {code}")   # all -> de

Full script: examples/normalization.py.

In an OVOS skill vs. standalone

The API is identical. Only where you get text and lang differs.

# Standalone language-routing utility
code, conf = extract_langcode(user_input, "en")

# Inside an OVOS skill, the utterance and its language come from the session
class MySkill(OVOSSkill):
    def handle_translate(self, message):
        utterance = message.data["utterance"]
        code, conf = extract_langcode(utterance, self.lang)
        ...

Coverage

Names can be written in 21 languages. Each carries a table of a few hundred target languages keyed by ISO 639 code.

The live list is always ovos_lang_parser.LANGS. Adding a language means dropping in one JSON file. See docs/coverage.md and docs/extending.md.

Documentation

• docs/api.md: full reference, confidence model, edge behavior
• docs/coverage.md: supported languages and the data model
• docs/extending.md: add a language wordlist
• examples/: runnable scripts for each use case

Related projects

• ovos-number-parser: numbers
• ovos-date-parser: dates and times
• ovos-color-parser: colors

License

Apache 2.0. See LICENSE.