Language Detector: Detect Text Language in Bulk (API)

This language detector runs language detection on texts, web pages and dataset fields in bulk. Paste sentences, page URLs or pick a dataset and get one row per item with the ISO 639-1 code and language name, a confidence score, the top 3 candidates, the writing script, an isReliable flag and the character count. Optional mixed-language detection splits a text into sentences and reports every language with its share. It runs offline inside the Actor: no API key, no external service, no model download at run time.

Pay only for items that got a language result. Empty, failed and skipped items are free.

What can this language identifier do

• Detect the language of 75 languages (English, Spanish, French, German, Portuguese, Russian, Chinese, Japanese, Korean, Arabic, Hindi and more) with the open-source Lingua library, which is strong on short texts.
• Bulk texts. One line per text, up to 50,000 items per run. One bad item never stops the run.
• Web pages. Give page URLs: the main text is extracted with trafilatura and detected. The page's own <html lang> attribute (or content-language / og:locale meta) is returned as a second signal, with htmlLangMatches telling you if the page declares the language it actually uses.
• Dataset fields. Pick the dataset of another Actor (reviews, comments, tweets, descriptions) and name the field, for example review.text.
• Top 3 candidates with confidence, so you can see close calls (Spanish vs Portuguese, Norwegian vs Danish).
• Script detection from Unicode ranges: Latin, Cyrillic, Greek, Arabic, Hebrew, Devanagari, CJK (Han), Kana (Japanese), Hangul, Thai and others.
• isReliable and low-confidence flag. Texts shorter than 20 characters always return lowConfidence: true and isReliable: false with a note. You choose the reliable threshold (default 0.7).
• Mixed languages (detectMixed). Sentence-by-sentence detection returns languages (language, sentences, characters, share) and isMixed when two or more languages each cover 15% or more.
• Restrict the choice. If you know the text is English, French or German, set restrictToLanguages to en, fr, de. This makes short texts much more accurate.

How to detect the language of a text in bulk

1. Paste your texts in Texts, one per line (the prefill has four sentences in four languages).
2. Run the Actor.
3. Download the dataset as JSON, CSV or Excel, or call it from the API. The language column holds the ISO 639-1 code.

How to detect the language of a web page

Put page URLs in Page URLs. You get the language of the readable text plus htmlLang. Public pages only: no login, no CAPTCHA solving, and login-walled social sites are refused. Pages that need JavaScript to show their text come back as empty with an explanation.

How to detect the language of a dataset field

Pick the Dataset and set Dataset field (default text). Every string in that field becomes one row. Dotted paths and list positions work (comments.0).

Input example

{
  "texts": [
    "The quick brown fox jumps over the lazy dog near the river bank.",
    "Hoy hemos publicado una nueva versión de la aplicación con muchas mejoras.",
    "Мы рады сообщить о запуске нашего нового сервиса для малого бизнеса.",
    "今日は天気がとても良いので、公園で友達と一緒にお昼ご飯を食べました。"
  ],
  "urls": ["https://en.wikipedia.org/wiki/Language"],
  "detectMixed": false,
  "restrictToLanguages": [],
  "maxItems": 1000
}

Output example

{
  "index": 2,
  "source": "text",
  "url": null,
  "textPreview": "Hoy hemos publicado una nueva versión de la aplicación con muchas mejoras.",
  "language": "es",
  "languageName": "Spanish",
  "confidence": 0.9986,
  "candidates": [
    {"language": "es", "name": "Spanish", "confidence": 0.9986},
    {"language": "ca", "name": "Catalan", "confidence": 0.0009},
    {"language": "pt", "name": "Portuguese", "confidence": 0.0003}
  ],
  "script": "Latin",
  "isReliable": true,
  "lowConfidence": false,
  "charCount": 75,
  "htmlLang": null,
  "htmlLangMatches": null,
  "status": "detected",
  "error": null,
  "note": null
}

With detectMixed each row also has isMixed and languages, for example [{"language":"en","name":"English","sentences":2,"charCount":96,"share":0.62}, ...]. A SUMMARY record in the key-value store counts results per language.

Use cases

• Route support tickets, reviews or comments to the right language team.
• Clean a scraped dataset: keep only English rows, or find the rows in an unexpected language.
• Audit a website: do pages declare the right lang attribute (htmlLangMatches)?
• Pick a translation or NLP model per text, or filter spam in the wrong language.

How much does it cost

Pay per event: text-detected = $0.0008 per text, page or dataset item with a language result ($0.80 per 1,000). Empty, failed, skipped and "undetected" (numbers or symbols only) rows are not charged. Platform compute usage is included in the price on paid plans; use Max items to cap spend.

FAQ

How accurate is the language detection? Lingua is among the most accurate open-source detectors for short and mixed text. Long, single-language text is almost always right. Very short texts (under 20 characters), names, code and closely related languages are guesses, which is why those rows are flagged lowConfidence and show the top 3 candidates.

Which languages are supported? The 75 languages of Lingua 2.x (all return an ISO 639-1 code). Languages outside that list are assigned to the closest known one, so use isReliable and the confidence to catch them.

What does confidence mean? The model's relative probability of the top language among the candidate languages (0 to 1). Restricting restrictToLanguages raises it, because there are fewer rivals.

Can it detect a text that mixes two languages? Yes, turn on detectMixed. It detects each sentence (very short sentences are merged with neighbours) and reports the share of each language. The main language still describes the whole text.

Does it detect the language of a web page? Yes, from the page's main text, and it also reports what the page declares in <html lang>. It does not render JavaScript, so pages that build their content in the browser may return empty.

Does it use an external API or my data? No. Detection runs inside the Actor. Page URLs are downloaded from the public internet; nothing is sent to a third-party language service.

Is the first run slow? The accurate mode loads its language models at start, which takes roughly 10-20 seconds; after that, thousands of texts take seconds. accuracy: fast starts quicker but is less accurate on short texts.

Limits

• 75 languages; no dialect or locale detection (en-GB vs en-US) beyond what a page declares.
• Texts are cut at 20,000 characters for detection (configurable). charCount shows the full length.
• Page URLs: public HTML pages up to 10 MB, no JavaScript rendering, no logins.

Related tools

• Doc to Text Extractor to get text out of PDFs and Office files before detecting the language.
• Keyword Suggestions to find what people search for in each language.
• SEO Audit to check page-level signals such as the lang attribute.