Guessing the encoding of text represented as byte[] in Java

Question

Given an array of bytes representing text in some unknown encoding (usually UTF-8 or ISO-8859-1, but not necessarily so), what is the best way to obtain a guess for the most likely encoding used (in Java)?

Worth noting:

• No additional meta-data is available. The byte array is literally the only available input.
• The detection algorithm will obviously not be 100 % correct. If the algorithm is correct in more than say 80 % of the cases that is good enough.

Solution 1

The following method solves the problem using juniversalchardet, which is a Java port of Mozilla's encoding detection library.

public static String guessEncoding(byte[] bytes) {
    String DEFAULT_ENCODING = "UTF-8";
    org.mozilla.universalchardet.UniversalDetector detector =
        new org.mozilla.universalchardet.UniversalDetector(null);
    detector.handleData(bytes, 0, bytes.length);
    detector.dataEnd();
    String encoding = detector.getDetectedCharset();
    detector.reset();
    if (encoding == null) {
        encoding = DEFAULT_ENCODING;
    }
    return encoding;
}

The code above has been tested and works as intented. Simply add juniversalchardet-1.0.3.jar to the classpath.

I've tested both juniversalchardet and jchardet. My general impression is that juniversalchardet provides the better detection accuracy and the nicer API of the two libraries.

Solution 2

There is also Apache Tika - a content analysis toolkit. It can guess the mime type, and it can guess the encoding. Usually the guess is correct with a very high probability.

Solution 3

Here's my favorite: https://github.com/codehaus/guessencoding

It works like this:

• If there's a UTF-8 or UTF-16 BOM, return that encoding.
• If none of the bytes have the high-order bit set, return ASCII (or you can force it to return a default 8-bit encoding instead).
• If there are bytes with the high bit set but they're arranged in the correct patterns for UTF-8, return UTF-8.
• Otherwise, return the platform default encoding (e.g., windows-1252 on an English-locale Windows system).

It may sound overly simplistic, but in my day-to-day work it's well over 90% accurate.

Solution 4

Chi's answer seems most promising for real use. I just want to add that, according to Joel Spolsky, Internet Explorer used a frequency-based guessing algorithm in its day:

http://www.joelonsoftware.com/articles/Unicode.html

Roughly speaking, all the assumed-to-be-text is copied, and parsed in every encoding imaginable. Whichever parse fits a language's average word (and letter?) frequency profile best, wins. I can not quickly see if jchardet uses the same kind of approach, so I thought I'd mention this just in case.

View more solutions

Comments

[knorv]

What about the cases where it is not UTF-8?

[knorv]

I kind of know how to use Google, but the question specifically asks for "what is the best way [..]". So which is best, icu4j, jchardet or some other library?

[knorv]

Please elaborate - why do you consider jchardet to be the best library around?

[coding_idiot]

my project requirement is if the data is not in utf8 (after detection) then convert it to utf8, how to do this ?

[coding_idiot]

@chi how to convert to utf8 if the encoding is not utf8.

[Brett Okken]

@coding_idiot use the "guessed" encoding to convert to a String then get utf-8 bytes: new String(bytes, guessedEncoding).getBytes("utf-8").

[james.garriss]

If it's not UTF-8, blindly calling it Latin-1 isn't a good idea. It would be better to use ICU, jchardet, or one of the other tools listed on this page to make an intelligent guess.

[Sxilderik]

Not very happy with this. See github.com/albfernandez/juniversalchardet/issues/22

[Aleksandr Erokhin]

juniversalchardet is also available in maven. groupId: com.googlecode.juniversalchardet, artifactId: juniversalchardet.