fstalign
is a tool for creating alignment between two sequences of tokens (here out referred to as “reference” and “hypothesis”). It has two key functions: computing word error rate (WER) and aligning NLP-formatted references with CTM hypotheses.
Due to its use of OpenFST and lazy algorithms for text-based alignment, fstalign
is efficient for calculating WER while also providing significant flexibility for different measurement features and error analysis.
Version 2.0 introduces two major changes:
- A new method to traverse the composition graph, which dramatically improves the overall speed, especially when the sequences are long contain many errors. We have files that took 25 minutes to align before that can now take about 7 seconds. This is especially noticeable with the adapted composition (the default).
- Some smarts were introduced when --use-case and --use-punctuation are enabled. Now, by default, punctuation symbols can only be substituted by other punctuation symbols (or deleted/inserted). Also, words that differ only by the first letter case will be preffered for substitution.
Here's an example of the 1.x behavior and the 2.0 version
==> v1.x sbs.txt <==
ref_token hyp_token IsErr Class Wer_Tag_Entities
Welcome Welcome ###322_###|
back back
to to
another another
episode episode ###323_###|
of of
Podcasts Podcast ERR ###324_###|
in and ERR
Color Color ###167_###|###325_###|
: of ERR
The the ERR
Podcast Podcast ###168_###|###326_###|
. .
I I
==> v2.0 sbs.txt <==
ref_token hyp_token IsErr Class Wer_Tag_Entities
Welcome Welcome ###322_###|
back back
to to
another another
episode episode ###323_###|
of of
Podcasts Podcast ERR ###324_###|
in and ERR
Color Color ###167_###|###325_###|
<ins> of ERR
: <del> ERR
The the ERR
Podcast Podcast ###168_###|###326_###|
The confusion between :
and of
is not longer allowed.
Also, here's how favoring or not the substitution based on case-insensitive comparison, while still counting it as an error, looks like:
==> v1.x sbs.txt <==
ref_token hyp_token IsErr Class Wer_Tag_Entities
shorten shorten ###801_###|
It's it's ERR
Berry Barry ERR ###785_###|###788_###|###802_###|
. .
Just Just
Yeah like ERR ###805_###|
. <del> ERR
Like <del> ERR
, <del> ERR
I I ###809_###|
have have
a a
nickname nickname
==> v2.0 sbs.txt <==
ref_token hyp_token IsErr Class Wer_Tag_Entities
It's it's ERR
Berry Barry ERR ###785_###|###788_###|###802_###|
. .
Just Just
Yeah <del> ERR ###805_###|
. <del> ERR
Like like ERR
, <del> ERR
I I ###809_###|
have have
a a
nickname nickname
Here, Like <-> like
substitution is favored. While this generally won't change the WER value itself (although it can), it will improve the timing alignments.
These behavior, as well as the beam size (that has a default value of 50.0) can be controlled with the following new parameters:
--disable-strict-punctuation
Disable strict punctuation alignment (which prevents punctuation aligning with words).
--disable-favored-subs Disable favored substitutions (which makes alignment favor substitutions between words which differ only by case).
--favored-sub-cost FLOAT Cost for favored substitutions (e.g., case diff). Default: 0.1
We use git submodules to manage third-party dependencies. Initialize and update submodules before proceeding to the main build steps.
git submodule update --init --recursive
This will pull the current dependencies:
- catch2 - for unit testing
- spdlog - for logging
- CLI11 - for CLI construction
- csv - for CTM and NLP input parsing
- jsoncpp - for JSON output construction
- strtk - for various string utilities
Additionally, we have dependencies outside of the third-party submodules:
- OpenFST - currently provided to the build system by settings the $OPENFST_ROOT environment variable or during the CMake command via
-DOPENFST_ROOT
.
The current build framework is CMake. Install CMake following the instructions here (https://cmake.org/install/).
To build fstalign, run:
mkdir build && cd build
cmake .. -DOPENFST_ROOT="<path to OpenFST>" -DDYNAMIC_OPENFST=ON
make
Note: -DDYNAMIC_OPENFST=ON
is needed if OpenFST at OPENFST_ROOT
is compiled as shared libraries. Otherwise static libraries are assumed.
Finally, tests can be run using:
make test
The fstalign docker image is hosted on Docker Hub and can be easily pulled and run:
docker pull revdotcom/fstalign
docker run --rm -it revdotcom/fstalign
See https://hub.docker.com/r/revdotcom/fstalign/tags for the available versions/tags to pull. If you desire to run the tool on local files you can mount local directories with the -v
flag of the docker run
command.
From inside the container:
/fstalign/build/fstalign --help
For development you can also build the docker image locally using:
docker build . -t fstalign-dev
For more information on how to use fstalign
see our documentation for more details.