texterrors 1.1.7

For WER

texterrors

texterrors scores ASR or transcription output against a reference and helps you inspect what went wrong.

It supports:

• WER and CER
• standard and character-aware alignment
• detailed aligned error reports
• colored output for inspecting alignments
• comparison of multiple hypothesis files against the same reference
• per-group metrics such as per-speaker WER
• keyword and OOV evaluation
• OOV-CER for targeted analysis of out-of-vocabulary words, as in this paper
• oracle WER
• simple entity accuracy
• per-entity diagnostic TSV output
• aggregate JSON output for scripting

Example of colored detailed output (--usecolor):

For more background on the motivation for the tool, see this post.

Changelog

1.1.7

• Restored the two-hypothesis per-utterance comparison view behind --side-by-side / -y.

Installing

Requires Python 3.9 or newer.

pip install texterrors

This installs both the Python package and the texterrors command-line tool.

Common usage

If your files are ark-like text with an utterance ID as the first field, use --isark.

Compute aggregate WER only:

texterrors --isark -s ref hyp

Write a detailed report to a file:

texterrors --isark --cer -c -o detailed_report.txt ref hyp

If you use --usecolor, view the output with less -R.

Compare several systems against the same reference:

texterrors --isark -s ref hyp1 hyp2 hyp3

This prints a comparison table with one row per hypothesis file.

Compare two systems in one aligned per-utterance block:

texterrors --isark -y ref hyp1 hyp2

Write aggregate-only JSON instead of the normal text report:

texterrors --isark --output-format json ref hyp

Measure simple entity accuracy:

texterrors --isark -w ref hyp

Write one TSV row per entity occurrence for diagnostics:

texterrors --isark --entity-details entity_details.tsv ref hyp

Input formats

By default, texterrors expects one reference line and one hypothesis line per utterance.

Useful input flags:

• --isark: each line starts with an utterance ID
• --isctm: input is CTM-like and includes timing fields

Output modes

The default output is a human-readable text report. Unless you pass --skip-detailed, it includes per-utterance aligned detail as well as overall summary statistics.

Useful output options:

• --skip-detailed: show only aggregate statistics
• --out, -o: write the text report to a file
• --side-by-side, -y: for exactly two hypothesis files, show reference and both hypotheses together per utterance
• --output-format json: write aggregate statistics and top-error summaries as JSON, without per-utterance detail
• --entity-details FILE: write a TSV with one row per simple-entity occurrence

--entity-details is meant for compact diagnostics. The TSV records the normalized compact hypothesis output aligned to the entity span, not the original surface form with exact spacing and casing.

Common analysis options

• --cer: compute CER in addition to WER
• --utt-group-map FILE: report metrics by group, for example by speaker
• --keywords-f FILE: restrict keyword precision and recall analysis to terms in a file
• --oov-list-f FILE: compute OOV-CER for words in a file
• --oracle-wer: pick the lowest-edit-distance hypothesis when multiple hypotheses are available per utterance
• --freq-sort: sort error summaries by frequency instead of raw count
• --num-top-errors N: control how many top insertions, deletions, and substitutions are shown

Entity scoring

--simple-entity-accuracy uses reference-side casing cues to identify likely entity terms and then scores whether those terms were recognized in the hypothesis.

Entity matching is case-insensitive and ignores whitespace, so examples like GenAI and Gen AI count as the same entity for this metric.

If you want to inspect misses, use --entity-details FILE. That TSV is useful for spotting whether an entity matched cleanly, was substituted, or was effectively deleted.

Python usage

If you want to use the library directly from Python:

from texterrors import align_texts

ref_aligned, hyp_aligned, cost = align_texts(
    ["speedbird", "eight", "six", "two"],
    ["hello", "speedbird", "six", "two"],
    use_chardiff=True,
)

Benchmarking

A small benchmark harness lives in benchmarks/alignment_benchmark.py.

Run it from the repo root:

.venv/bin/python benchmarks/alignment_benchmark.py --repeat 7

You can also point it at other ark-like files:

.venv/bin/python benchmarks/alignment_benchmark.py --ref my_ref.txt --hyp my_hyp.txt

Development install

If you want to build the extension locally:

uv venv
env UV_CACHE_DIR=/tmp/uv-cache uv pip install --python .venv/bin/python -r requirements.txt
.venv/bin/cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DPython_EXECUTABLE=$PWD/.venv/bin/python -Dnanobind_DIR=$PWD/.venv/lib/python3.12/site-packages/nanobind/cmake
.venv/bin/cmake --build build --config Release
.venv/bin/cmake --install build --config Release --prefix $PWD

Note on --use-chardiff

--use-chardiff enables character-aware alignment. This often gives more intuitive alignments when words are similar, but it can also make WER slightly higher than standard token-only alignment.

If you want behavior closer to a plain token-alignment scorer, leave --use-chardiff off.

For example, a plain token alignment might force a one-to-one mapping:

test	sentence	okay	words	ending	now
test	a	sentenc	ok	endin	now

Character-aware alignment may instead align it like this:

test	-	sentence	okay	words	ending	now
test	a	sentenc	ok	-	endin	now

That can increase WER because it exposes an insertion and deletion that the token-only alignment hides.