SubER is an automatic, reference-based, segmentation- and timing-aware edit distance metric to measure quality of subtitle files. For a detailed description of the metric and a human post-editing evaluation we refer to our IWSLT 2022 paper. In addition to the SubER metric, this scoring tool calculates a wide range of established speech recognition and machine translation metrics (WER, BLEU, TER, chrF) directly on subtitle files.
pip install subtitle-edit-ratewill install the suber command line tool.
Alternatively, check out this git repository and run the contained suber module with python -m suber.
Currently, we expect subtitle files to come in SubRip text (SRT) format. Given a human reference subtitle file reference.srt and a hypothesis file hypothesis.srt (typically the output of an automatic subtitling system) the SubER score can be calculated by running:
$ suber -H hypothesis.srt -R reference.srt
{
"SubER": 19.048
}The SubER score is printed to stdout in json format. As SubER is an edit rate, lower scores are better. As a rough rule of thumb from our experience, a score lower than 20(%) is very good quality while a score above 40 to 50(%) is bad.
Make sure that there is no constant time offset between the timestamps in hypothesis and reference as this will lead to incorrect scores.
Also, note that <i>, <b> and <u> formatting tags are ignored if present in the files. All other formatting must be removed from the files before scoring for accurate results.
The main SubER metric is computed on normalized text, which means case-insensitive and without taking punctuation into account, as we observe higher correlation with human judgements and post-edit effort in this setting. We provide an implementation of a case-sensitive variant which also uses a tokenizer to take punctuation into account as separate tokens which you can use "at your own risk" or to reassess our findings. For this, add --metrics SubER-cased to the command above. Please do not report results using this variant as "SubER" unless explicitly mentioning the punctuation-/case-sensitivity.
The SubER tool supports computing the following other metrics directly on subtitle files:
- word error rate (WER)
- bilingual evaluation understudy (BLEU)
- translation edit rate (TER)
- character n-gram F score (chrF)
- character error rate (CER)
BLEU, TER and chrF calculations are done using SacreBLEU with default settings. WER is computed with JiWER on normalized text (lower-cased, punctuation removed).
Assuming hypothesis.srt and reference.srt are parallel, i.e. they contain the same number of subtitles and the contents of the n-th subtitle in both files corresponds to each other, the above-mentioned metrics can be computed by running:
$ suber -H hypothesis.srt -R reference.srt --metrics WER BLEU TER chrF CER
{
"WER": 23.529,
"BLEU": 39.774,
"TER": 23.529,
"chrF": 68.402,
"CER": 17.857
}In this mode, the text from each parallel subtitle pair is considered to be a sentence pair.
In the general case, subtitle files for the same video can have different numbers of subtitles with different time stamps. All metrics - except SubER - usually require to be calculated on parallel segments. To apply these metrics to general subtitle files, the hypothesis file has to be re-segmented to correspond to the reference subtitles. The SubER tool implements two options:
- alignment by minimizing Levenshtein distance (Matusov et al.)
- time alignment method from Cherry et al.
See our paper for further details.
To use the Levenshtein method add an AS- prefix to the metric name, e.g.:
suber -H hypothesis.srt -R reference.srt --metrics AS-BLEUThe AS- prefix terminology is taken from Matusov et al. and stands for "automatic segmentation".
To use the time-alignment method instead, add a t- prefix. This works for all metrics (except for SubER itself which does not require re-segmentation). In particular, we implement t-BLEU from Cherry et al.. We encode the segmentation method (or lack thereof) in the metric name to explicitly distinguish the different resulting metric scores!
To inspect the re-segmentation applied to the hypothesis you can use the align_hyp_to_ref.py tool (run python -m suber.tools.align_hyp_to_ref -h for help).
In case of Levenshtein alignment, there is also the option to give a plain file as the reference. This can be used to provide sentences instead of subtitles as reference segments (each line will be considered a segment):
suber -H hypothesis.srt -R reference.txt --reference-format plain --metrics AS-TERWe provide a simple tool to extract sentences from SRT files based on punctuation:
python -m suber.tools.srt_to_plain -i reference.srt -o reference.txt --sentence-segmentationIt can be used to create the plain sentence-level reference reference.txt for the scoring command above.
The line breaks present in the subtitle files can be included into the text segments to be scored as <eol> (end of line) and <eob> (end of block) tokens. For example:
636
00:50:52,200 -> 00:50:57,120
Ladies and gentlemen,
the dance is about to begin.
would be represented as
Ladies and gentlemen, <eol> the dance is about to begin. <eob>
To do so, add a -seg ("segmentation-aware") postfix to the metric name, e.g. BLEU-seg, AS-TER-seg or t-WER-seg. Character-level metrics (chrF and CER) do not support this as it is not obvious how to count character edits for <eol> tokens.
As a special case, we implement TER-br from Karakanta et al.. It is similar to TER-seg, but all (real) words are replaced by a mask token. This would convert the sentence from the example above to:
<mask> <mask> <mask> <eol> <mask> <mask> <mask> <mask> <mask> <mask> <eob>
Note, that also TER-br has variants for computing it on existing parallel segments (TER-br) or on re-aligned segments (AS-TER-br/t-TER-br). Re-segmentation happens before masking.
If you run into an issue, have a feature request or have questions about the usage or the implementation of SubER, please do not hesitate to open an issue or a thread under "discussions". Pull requests are welcome too, of course!
Things I'm already considering to add in future versions:
- support for other subtitling formats than SRT
- a verbose output that explains the SubER score (list of edit operations)
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent