Textstat-rs

Rust port of the textstat textual analysis library, with Python bindings. Drop-in replacements for 11 key metrics in British and US English.

99.98% exact output match, with at least 2.4x-7x speedup, with 2 most sped-up metrics at ~14x and ~79x based on benchmarking scripts included in repo, running on Wikipedia dataset as a varied, modern text corpus.

Installation

pip install textstat-rs

Pre-release versions need pip install --pre textstat-rs.

Requires Python 3.10+. Wheels are built for Linux (x86_64, x86, aarch64, armv7; glibc and musl), macOS (x86_64, arm64) and Windows (x64, arm64). They are abi3 wheels, so one wheel per platform covers every supported Python version. On other platforms pip falls back to the sdist, which needs a Rust toolchain to build.

Usage

Import name is textstat_rs; every metric takes the text as its first argument.

import textstat_rs

text = "The cat sat on the mat. It was a sunny day."

textstat_rs.flesch_reading_ease(text)   # 108.96159090909092
textstat_rs.flesch_kincaid_grade(text)  # -0.5722727272727273
textstat_rs.text_standard(text)         # '0th and 1st grade'

The exposed metrics match textstat's signatures, so an existing import can be swapped in place:

import textstat_rs as textstat

Optional arguments are the same as the Python library, other than the exceptions noted below:

textstat_rs.gunning_fog(text, syllable_threshold=3)
textstat_rs.linsear_write_formula(text, strict_lower=False, strict_upper=True)
textstat_rs.reading_time(text, ms_per_char=14.69)

The counting helpers the formulas are built on are exposed too: syllable_count, sentence_count, lexicon_count, char_count, letter_count, polysyllabcount, miniword_count and difficult_words.

Language

British and American English are both supported. The default is en_US, matching textstat.

Set it once, the way textstat does:

textstat_rs.set_lang("en_GB")
textstat_rs.get_lang()                  # 'en_GB'
textstat_rs.syllable_count("colourful") # 3   (2 under en_US)

supported_langs() returns canonical tags, to allow enumerating locales rather than hardcoding them:

textstat_rs.supported_langs()          # ['en_US', 'en_GB']

You can also set the locale per call. It is keyword-only, and overrides over the default:

textstat_rs.flesch_reading_ease(text, lang="en_GB")

lang= is offered only where the locale can change the answer. It reaches exactly one decision — which hyphenation dictionary spells out syllables for words missing from CMUdict — so char_count, lexicon_count, sentence_count, letter_count, miniword_count, reading_time, mcalpine_eflaw, coleman_liau_index and automated_readability_index do not take it, because it would do nothing.

Tags are resolved the way pyphen does — lowercased, - normalised to _, then trailing subtags dropped until something matches — so en-GB, EN_GB and en_US_posix all work.

Four differences from textstat worth knowing:

• Only English is supported. textstat would accept "fr" and reach for French coefficients and a French dictionary. textstat-rs vendors only the two English hyphenation dictionaries, so anything that doesn't resolve to en_US or en_GB raises ValueError.
• Bad locales fail immediately. textstat's set_lang is a bare assignment, so a typo is accepted and only surfaces later as a KeyError from inside pyphen — and only once some word actually misses CMUdict, which for a short text may be never.
• Bare "en" means British. pyphen registers a short name for the first matching file in sorted order, and hyph_en_GB.dic sorts before hyph_en_US.dic. Kept deliberately, so en and en_AU get the British dictionary. The default when nothing is set is still en_US.
• lang= is honoured. textstat still accepts syllable_count(text, lang=...) but warns and discards it. Here it works.

Python Comparison

Functions

Exposed Metrics

Parity

Scores are compared against textstat over 1000 Wikipedia articles (up to 5000 chars each), across en_US and en_GB: 99.98% of 22000 comparisons match exactly, and the largest disagreement anywhere is 0.049566.

Metrics returning a grade band rather than a number are compared as exact strings: text_standard matches exactly 1000/1000 in en_US, 1000/1000 in en_GB.

Generated from dc1a373 on 2026-08-24, python 3.12.4 vs textstat 0.7.13.

Performance

One call per metric over 100 concatenated Wikipedia articles (0.46 MB), median of 5. Both libraries are warmed first, so these are steady-state numbers with lazy resource loading excluded.

On top of that, the first call in a fresh process pays a one-off load of the syllable and word-list resources: up to 34 ms for textstat-rs against 361 ms for textstat.

Generated from dc1a373 on 2026-08-24, python 3.12.4 vs textstat 0.7.13.

Third-party data

The scoring resources are compiled into the extension by include_str!, so they ship inside every wheel. Their notices ship with them, under textstat_rs-<version>.dist-info/licenses/.

Both .dic files are vendored byte-for-byte from pyphen, and the notice files keep their upstream names so they can be diffed against it directly. The pyphen-LICENSE.txt note covers pyphen itself (GPL 2.0+/LGPL 2.1+/MPL 1.1 tri-licence). None of pyphen's code is used here, only the two dictionaries, which carry the BSD-style terms above.

Only the data is third-party. The Rust and Python code in this repository is MIT, per LICENSE.