About Language Detector

Detect text language automatically with confidence scores. Part of the DevTools Surf developer suite. Browse more tools in the Text / String collection.

Use Cases

• Automatically route customer support messages to the correct language team without manual inspection.
• Label a multilingual text dataset before training a language-specific NLP model.
• Verify that localized content is actually in the intended language before publishing.
• Identify the language of scraped web content before applying language-specific processing.

Tips

• Provide at least 50-100 characters of text for reliable detection — single words are ambiguous across similar languages (Dutch vs. Afrikaans, Serbian vs. Croatian).
• Check the confidence score alongside the detected language — low confidence (below 70%) signals short or mixed-language input that may be misclassified.
• Use the batch mode to classify a list of texts in one operation for dataset labeling or routing pipelines.

Fun Facts

• Compact Language Detector (CLD3), used in Google Chrome since 2016, identifies 107 languages using a neural network trained on web-crawled text. It runs locally in the browser without a network call.
• The Naïve Bayes language classifier, still competitive despite its 1960s origins, achieves 99%+ accuracy on long texts because language identity is highly determined by character n-gram frequencies.
• The longest-documented gap in language detection accuracy is between written Norwegian Bokmål and Danish — they share so much vocabulary that automated classifiers misidentify them at rates as high as 20% on short texts.

FAQ

How many languages does it support?

80+ languages including all major European, Asian, and Semitic languages. Less commonly digitized languages have lower accuracy due to limited training data.

Does it work on code-switched text (multiple languages mixed)?

It returns the dominant language with a confidence score. Code-switching detection (identifying language segments within a text) is a separate, harder problem not covered here.

What's the minimum text length for reliable detection?

50+ characters gives 95%+ accuracy for most language pairs. Under 20 characters, accuracy drops significantly. Single-word detection is unreliable for closely related language pairs.