finos / greenkey-asrtoolkit Goto Github PK

Describe the bug
Given a list such as "a,b" - clean_formatting currently just removes the comma. We should replace this with a space in the event bad data has entered a data processing pipeline.

To Reproduce
Steps to reproduce the behavior:

1. echo a,b > foo.txt
2. clean_formatting foo.txt
3. cat foo.txt

ab

Expected behavior

a b

Describe the bug
Occasionally, <#s> can be output by some decoders to indicate silence

Expected behavior
We should amend the regex matches for noise tags to incorporate this specific tag.

Additional context
This may be somewhat unclear if we group more tagged non speech events in the same variable, so we may want to rename the variable.

Given the lack of activity on this repo, I propose archiving it.

If and when activity can be restored, the repository can be easily unarchived, by either:

• reopening this issue
• emailing [email protected]

We will wait until Friday the 25th of November to collect feedback from the community, then we will proceed with the archival.

Archiving the repo will consist of:

• Adding a note at the top of the repo home - https://github.com/finos/greenkey-asrtoolkit
• make the repository read-only, though it will still be public and downloadable

Subtitle files in SRT format should be supported by this tool

With this file as test.txt

okay
yeah
okay

I get this as an output

okayyeahokay

The current audio corpora prep seems to only work on SPH files. In addition, the current description says this:

 Note that filenames with hyphens will be sanitized to underscores and that audio files will be forced to single channel, 16 kHz, signed PCM format. If two channels are present, only the first will be used.

Many corpora come in WAV files instead of SPH files, and many also have two unmixed channels that need to be mixed to properly account for all audio.

Is your feature request related to a problem? Please describe.
Noise tags [noise] or <noise> are not removed by the text clean up routines.

Describe the solution you'd like
These should be remove using a regex that matches A-Za-z words in brackets or angle brackets

In some cases, it's desirable to simply extract a readable transcript out of JSON. For example, a text representation or well-formatted subtitles. In these cases, it would be useful to have convert_transcript output the formatted / punctuated transcript instead of the unformatted one.

As agreed to in our last PMC meeting, please move the project roadmap from confluence to this project, wherever it fits best (I suggest the Wiki, with a link from the README).

https://finosfoundation.atlassian.net/wiki/spaces/VOICE/pages/906133835/GreenKey+ASR+Toolkit+Roadmap

What kind of inputs does the wer tool support? Could you include some examples in the documentation?

Describe the bug
Line 58 of align_json.py: https://github.com/finos/greenkey-asrtoolkit/blob/master/asrtoolkit/align_json.py

Should this be "align_json(", not "align_gk_json("? I don't see align_gk_json anywhere else in the repo.

To Reproduce
Steps to reproduce the behavior:

1. Go to '...'
2. Click on '....'
3. Execute '....'
4. See error

Expected behavior
A clear and concise description of what you expected to happen.

Screenshots
If applicable, add screenshots to help explain your problem.

Desktop (please complete the following information):

• OS: [e.g. iOS]
• Browser [e.g. chrome, safari]
• Version [e.g. 22]
• Python version
• Using docker or pip package?

Additional context
Add any other context about the problem here.

Is your feature request related to a problem? Please describe.
GK JSON schema has been better defined since asrtoolkit first output to it.

Describe the solution you'd like
They should be floats

Describe alternatives you've considered
We could maintain both or switch to strings

Additional context
GK schema last edited by @burrows

Hi I use Python 3.5.3 :: Anaconda custom (x86_64) and when I ran the command wer answer.txt stt.txt the result was wer 400%

My another coworker tried this with same files, same comand, the python version was 3.6.8 but he got the result seemed normal like wer 49.485 % . His computer is Ubuntu and I use Mac OS

Other coworkers (One Ubuntu, one Windows) also tried and they got the wer 400% result like me so can you help me to figure out the reason that the calculation result is different from me and the other person even though using same files and same command?

Thank you!!!

Is your feature request related to a problem? Please describe.
We need to provide clean HTML output that can be served in front end applications

Describe the solution you'd like
We should be able to return html output as a string or as a file

Many use cases directly call sanitize hyphens. However, there's no choice in these cases to disable the warning message.

Need to support VTT format for file conversion

Frequently takes a long time and much memory to compute the word error rate

Describe the solution you'd like
validate_stm should be a command line tool

Describe alternatives you've considered
Users can invoke this by writing python code

Additional context
Requested by users of asrtoolkit

Describe the bug
The wer script accepts pretty much anything as input and will spit out a random WER. This conflates a problem with the wer script and a simple filename error. I would suggest adding

To Reproduce
Steps to reproduce the behavior:

1. Run the wer script with any invalid input
2. Observe a lack of error and an output WER

> wer asdf asd
WER: 0.000%
> wer ^ ^
WER: 0.000%
> wer aaaaaaaaaaaaaaa eeeeeeeeeeeeeeeeeee
WER: 100.000%
> wer --ignore-nsns I-like to-eat-garbage
WER: 150.000%
> wer seg_data/gkt_corpora_earnings_test_AAPL2017q1TranscriptMturkTest_seg_0001. seg_data/gkt_corpora_earnings_test_AAPL2017q1TranscriptMturkTest_seg_0001.st
WER: 11.111%

Expected behavior
I would expect the wer script to error when given inputs that aren't valid ( ie. the expect file format, imaginary files, etc. )

Desktop (please complete the following information):

• OS: Ubuntu 16.04.5
• Python version 3.5.2+

Is your feature request related to a problem? Please describe.
Phone numbers in transcripts are mapped to long series of numerals

Describe the solution you'd like
they should be mapped to spelled out single digits

Examples:

1-317-222-222 should map to the series of numbers 'one', 'three', 'one', 'seven', etc.

Describe the solution you'd like
The quantity "$6.5B" should map to "six point five billion" in the un-formatted text

Testing coverage gap here

The excel spreadsheet reader in asrtoolkit (as written) requires pandas, which is dependent upon numpy. The latter module was removed on 1/28/19 from requirements.txt to address test failures on circleci (cause undetermined). This script is rarely used internally, and does not affect the rest of the codebase but it is used by clients.

To Do:

fix the excel spreadsheet reader in asrtoolkit using a library other than pandas. One option would be to use csv with the excel format option

Basic json formatting should be supported