Language identification

POST /v1/language-idPOST /v1/language-id/batchbilled per text

Identifies 608 languages and holds up on short, messy user input. Powered by lid-neural-608 and lid-lite-608.

r = call("/v1/language-id", {"text": "Mo fẹ́ lọ sí ọjà"})
print(r["language"], r["confidence"])   # yor 0.97

Request

Field	Default	Notes
text	required	Up to 8,000 characters.
model	lid-neural-608	lid-neural-608 is the most accurate; lid-lite-608 (fastText) is the fastest.
mode	traffic	traffic is for user input, chat and routing: on short text it leans towards the languages people actually use. coverage treats every language equally, for building corpora.
top_k	3	1 to 25 runner-up languages in top.

Response

{
  "language": "yor", "script": "Latn", "label": "yor_Latn", "language_name": "Yoruba",
  "confidence": 0.97,
  "top": [{"label": "yor_Latn", "language": "yor", "score": 0.97},
          {"label": "ibo_Latn", "language": "ibo", "score": 0.01}],
  "warnings": [],
  "model": "lid-neural-608", "mode": "traffic", "usage": {"texts": 1}
}

• language is ISO 639-3, and label adds the script (yor_Latn, srp_Cyrl).
• zxx_Zxxx means no language was found: the text is code, numbers or symbols. A warning says so.
• language_name is filled in for the 25 question-generation languages, and is the code otherwise.

Batch

POST /v1/language-id/batch with "texts": [...] (up to 256) returns {"results": [...]} in the same order.