`py3langid_rs`

A high-performance, pure Rust implementation of language identification, ported from the Python library py3langid.

[!NOTE] This implementation contains minimum functionalities. It lacks server, probability normalization, language subset.

Usage

Add to your project:

cargo add py3langid_rs

Example usage:

use py3langid_rs::LanguageIdentifier; 

fn main() {
    let li = LanguageIdentifier::new();
    println("{:?}", li.classify("This text is in English."));
}

Code above should print ("en", -56.77429).

Performance

AMD Ryzen 9 5950X, rustc 1.84.0 (9fc6b4312 2025-01-07), Ubuntu 22.04 in WSL 2.3.26.0, Windows 11 23H2.

Implementation	Lang	Slope	Median	Mean	Std. Dev.	Speed up (Slope)
`py3langid_rs`	en	29.153 µs	29.158 µs	29.169 µs	213.92 ns	20.954x
`py3langid`	en	610.884 µs	658.544 µs	610.884 µs	161.042 µs	1.0x
`py3langid_rs`	zh	14.521 µs	14.476 µs	14.502 µs	56.782 ns	31.296x
`py3langid`	zh	454.454 µs	489.616 µs	454.454 µs	75.018 µs	1.0x
`py3langid_rs`	jp	20.472 µs	20.415 µs	20.464 µs	149.43 ns	33.969x
`py3langid`	jp	695.421 µs	747.794 µs	695.421 µs	114.144 µs	1.0x

Using custom model

The converted model is uploaded to git, thus normally you don't have to do this. Only do this when there's a model update in the upstream, or you have a customly trained model.

There's no easy way to directly load the original pickle. Thus, we must convert the pickle first.

Set up environment

I'm using uv here due to it's super fast speed, you can also use other package managers.

uv venv
uv sync

Run conversion script

uv run convert_pkl.py path/to/your/model.plzma path/to/output/folder

This would automatically create/overwrite file model.bin in the output folder. Then in rust, load like this:

use py3langid_rs::LanguageIdentifier; 

fn main() {
    let li = LanguageIdentifier::from_lzma_file("path/to/output/folder/model.bin").unwrap();
    println("{:?}", li.classify("This text is in English."));
}