Efficient language detector (Nito-ELD or ELD) is a fast and accurate natural language detection software, written 100% in PHP, with a speed comparable to fast C++ compiled detectors, and accuracy rivaling the best detectors to date.

It has no dependencies, easy installation, all it's needed is PHP with the mb extension.
ELD scales perfectly with database size.
ELD is also available (outdated versions) in Javascript and Python.

1. Installation
2. How to use
3. Benchmarks
4. Databases
5. Testing
6. Languages

Changes from ELD v2 to v3:

• detect()->language now returns string 'und' for undetermined instead of NULL
• Databases are not compatible, and bigger, medium v2 ≈ small v3
• dynamicLangSubset() function is removed
• Function cleanText() is now named enableTextCleanup()

$ composer require nitotm/efficient-language-detector

• --prefer-dist will omit tests/, misc/ & benchmark/, or use --prefer-source to include everything
• Install nitotm/efficient-language-detector:dev-main to try the last unstable changes
• Alternatively, download / clone the files can work just fine.
  (Only small DB install under construction)

• ELD has database execution Modes: array, string, bytes, disk
• ELD database Sizes: small, medium, large, extralarge
• Modes string, bytes, disk only ship with database sizes: small & extralarge, but others can be built with BlobDataBuilder()

Low memory Modes
For Modes string, bytes and disk, all size databases can run under 128MB
disk mode with extralarge size, can run with 8MB setting memory_limit=8M, as it only uses 0.5MB

Fastest Mode: Array
For array Mode it is recommended to use OPcache, specially for the larger databases to reduce load times
We need to set opcache.interned_strings_buffer, opcache.memory_consumption high enough for each database
Recommended value in parentheses. Check Databases for more info.

detect() expects a UTF-8 string and returns an object with a language property, containing an ISO 639-1 code (or other selected scheme), or 'und' for undetermined language.

// require_once 'manual_loader.php'; To load ELD without autoloader. Update path.
use Nitotm\Eld\{LanguageDetector, EldDataFile, EldScheme, EldMode};

// LanguageDetector(databaseFile: ?string, outputFormat: ?string, mode: string)
$eld = new LanguageDetector(EldDataFile::SMALL, EldScheme::ISO639_1, EldMode::MODE_ARRAY);
// Database files: 'small', 'medium', 'large', 'extralarge'. Check memory requirements
// Database modes: 'array', 'string', 'bytes', 'disk'.
// Schemes: 'ISO639_1', 'ISO639_2T', 'ISO639_1_BCP47', 'ISO639_2T_BCP47' and 'FULL_TEXT'
// Modes 'string', 'bytes' & 'disk' only ship with DB sizes 'small' & 'extralarge'
// Constants are not mandatory, LanguageDetector('small'); will also work

$eld->detect('Hola, cómo te llamas?');
// object( language => string, scores() => array<string, float>, isReliable() => bool )
// ( language => 'es', scores() => ['es' => 0.25, 'nl' => 0.05], isReliable() => true )

$eld->detect('Hola, cómo te llamas?')->language;
// 'es'

Calling langSubset() once, will set the subset.

• In array mode the first call takes longer as it creates a new database, if save enabled (default), it will be loaded next time we make the same subset.
• In modes string, bytes & disk, a "virtual" subset is created instantly, detect() will just remove unwanted languages before returning results.
• To load a subset with 0 overhead, we can feed the returned file by langSubset() in array mode, when creating the instance LanguageDetector(file)
  • To make use of pre-built subsets in modes string, bytes & disk, getting lower memory usage and increased speed, it is possible by manually converting an array database, using BlobDataBuilder()
• Check available Languages below.

// It accepts any ISO codes.
// langSubset(languages: [], save: true, encode: true); Will return subset file name if saved
$eld->langSubset(['en', 'es', 'fr', 'it', 'nl', 'de']);
// Object ( success => bool, languages => ?array, error => ?string, file => ?string )
// ( success => true, languages => ['en', 'es'...], error => NULL, file => 'small_6_mfss...' )

// to remove the subset
$eld->langSubset();

// Load pre-saved subset directly, just like a default database
$eld_subset = new Nitotm\Eld\LanguageDetector('small_6_mfss5z1t');

// Build a binary database for modes 'string', 'bytes' & 'disk', from any 'array' database
// Memory requirements for 'array' database input apply 
$eldBuilder = new BlobDataBuilder('large'); // or subset 'small_6_mfss5z1t'
// Create subset directly: new BlobDataBuilder('extralarge', ['en', 'es', 'de', 'it']);
$eldBuilder->buildDatabase();

// if enableTextCleanup(True), detect() removes Urls, .com domains, emails, alphanumerical...
// Not recommended, as urls & domains contain hints of a language, which might help accuracy
$eld->enableTextCleanup(true); // Default is false

// If needed, we can get info of the ELD instance: languages, database type, etc.
$eld->info();

// Change output scheme on demand
// 'ISO639_1', 'ISO639_2T', 'ISO639_1_BCP47', 'ISO639_2T_BCP47', 'FULL_TEXT'
$eld->setOutputScheme('ISO639_2T'); // returns bool true on success

There is a CLI wrapper (BETA version)
>./bin/eld --help on Linux.
>php bin/eld --help on Windows.

I compared ELD with a different variety of detectors, as there are not many in PHP.

Benchmarks:

• Tatoeba: 20MB, short sentences from Tatoeba, 50 languages supported by all contenders, up to 10k lines each.

• For Tatoeba, I limited all detectors to the 50 languages subset, making the comparison as fair as possible.
• Also, Tatoeba is not part of ELD training dataset (nor tuning), but it is for fasttext

• ELD Test: 10MB, sentences from the 60 languages supported by ELD, 1000 lines each. Extracted from the 60GB of ELD training data.
• Sentences: 8MB, sentences from Lingua benchmark, minus unsupported languages and Yoruba which had broken characters.
• Word pairs 1.5MB, and Single words 870KB, also from Lingua, same 53 languages.

• Lingua participates with 54 languages, Franc with 58, patrickschur with 54.
• fasttext does not have a built-in subset option, so to show its accuracy and speed potential I made two benchmarks, fasttext-all not being limited by any subset at any test
• ^* Google's CLD2 also lacks subset option, and it's difficult to make a subset even with its option bestEffort = True, as usually returns only one language, so it has a comparative disadvantage.
• Time is normalized: (total lines * time) / processed lines

Special mention to 'disk' mode, while slower, is the fastest uncached load & detect for up to a 100 detections.
Modes 'bytes' and 'string' are very similar, they differ on how they are load, and are just 2x slower than Array
Modes 'string', 'bytes' & 'disk' only ship with DB sizes 'small' & 'extralarge'

• * I recommend using more than enough interned_strings_buffer as buffers overflow error might delay server response.
  To use all databases opcache.interned_strings_buffer should be a minimum of 160MB (170MB).
• When choosing the amount of memory keep in mind opcache.memory_consumption includes opcache.interned_strings_buffer.
  • If OPcache memory is 230MB, interned_strings is 32MB, and medium DB is 69MB cached, we have a total of (230 -32 -69) = 129MB of OPcache for everything else.
• Also, if you are going to use a subset of languages in addition to the main database, or multiple subsets, increase opcache.memory accordingly if you want them to be loaded instantly. To cache all default databases comfortably you would want to set it at 1200MB.

Default composer install might not include these files. Use --prefer-source to include them.

• For dev environment with composer "autoload-dev" (root only), the following will execute the tests

new Nitotm\Eld\Tests\TestsAutoload();

• Or, you can also run the tests executing the following file:

$ php efficient-language-detector/tests/tests.php # Update path

• To run the accuracy benchmarks run the benchmark/bench.php file.

• These are the ISO 639-1 codes that include the 60 languages. Plus 'und' for undetermined
  It is the default ELD language scheme. outputScheme: 'ISO639_1'

am, ar, az, be, bg, bn, ca, cs, da, de, el, en, es, et, eu, fa, fi, fr, gu, he, hi, hr, hu, hy, is, it, ja, ka, kn, ko, ku, lo, lt, lv, ml, mr, ms, nl, no, or, pa, pl, pt, ro, ru, sk, sl, sq, sr, sv, ta, te, th, tl, tr, uk, ur, vi, yo, zh

• These are the 60 supported languages for Nito-ELD. outputScheme: 'FULL_TEXT'

Amharic, Arabic, Azerbaijani (Latin), Belarusian, Bulgarian, Bengali, Catalan, Czech, Danish, German, Greek, English, Spanish, Estonian, Basque, Persian, Finnish, French, Gujarati, Hebrew, Hindi, Croatian, Hungarian, Armenian, Icelandic, Italian, Japanese, Georgian, Kannada, Korean, Kurdish (Arabic), Lao, Lithuanian, Latvian, Malayalam, Marathi, Malay (Latin), Dutch, Norwegian, Oriya, Punjabi, Polish, Portuguese, Romanian, Russian, Slovak, Slovene, Albanian, Serbian (Cyrillic), Swedish, Tamil, Telugu, Thai, Tagalog, Turkish, Ukrainian, Urdu, Vietnamese, Yoruba, Chinese

• ISO 639-1 codes with IETF BCP 47 script name tag. outputScheme: 'ISO639_1_BCP47'

am, ar, az-Latn, be, bg, bn, ca, cs, da, de, el, en, es, et, eu, fa, fi, fr, gu, he, hi, hr, hu, hy, is, it, ja, ka, kn, ko, ku-Arab, lo, lt, lv, ml, mr, ms-Latn, nl, no, or, pa, pl, pt, ro, ru, sk, sl, sq, sr-Cyrl, sv, ta, te, th, tl, tr, uk, ur, vi, yo, zh

• ISO 639-2/T codes (which are also valid 639-3) outputScheme: 'ISO639_2T'. Also available with BCP 47 ISO639_2T_BCP47

amh, ara, aze, bel, bul, ben, cat, ces, dan, deu, ell, eng, spa, est, eus, fas, fin, fra, guj, heb, hin, hrv, hun, hye, isl, ita, jpn, kat, kan, kor, kur, lao, lit, lav, mal, mar, msa, nld, nor, ori, pan, pol, por, ron, rus, slk, slv, sqi, srp, swe, tam, tel, tha, tgl, tur, ukr, urd, vie, yor, zho

If you wish to donate for open source improvements, hire me for private modifications, request alternative dataset training, or contact me, please use the following link: https://linktr.ee/nitotm