asrtoolkit

GreenKey Automatic Speech Recognition (ASR) Toolkit

---

The GreenKey ASRToolkit provides tools for file conversion and ASR corpora organization. These are intended to simplify the workflow for building, customizing, and analyzing ASR models, useful for scientists, engineers, and other technologists in speech recognition.

File formats supported

File formats have format-specific handlers in asrtoolkit/data_handlers. The scripts convert_transcript and wer support stm, srt, vtt, txt, and GreenKey json formatted transcripts. A custom html format is also available, though this should not be considered a stable format for long term storage as it is subject to change without notice.

convert_transcript

usage: convert_transcript [-h] input_file output_file

convert a single transcript from one text file format to another

positional arguments:
  input_file   input file
  output_file  output file

optional arguments:
  -h, --help   show this help message and exit

This tool allows for easy conversion among file formats listed above.

Note: Attributes of a segment object not present in a parsed file retain their default values

• For example, a segment object is created for each line of an STM line
• each is initialized with the following default values which are not encoded in STM files: formatted_text=''; confidence=1.0

wer

usage: wer [-h] [--char-level] [--ignore-nsns]
           reference_file transcript_file

Compares a reference and transcript file and calculates word error rate (WER)
between these two files

positional arguments:
  reference_file   reference "truth" file
  transcript_file  transcript possibly containing errors

optional arguments:
  -h, --help       show this help message and exit
  --char-level     calculate character error rate instead of word error rate
  --ignore-nsns    ignore non silence noises like um, uh, etc.

This tool allows for easy comparison of reference and hypothesis transcripts in any format listed above.

clean_formatting

usage: clean_formatting.py [-h] files [files ...]

cleans input *.txt files and outputs *_cleaned.txt

positional arguments:
  files       list of input files

optional arguments:
  -h, --help  show this help message and exit

This script standardizes how abbreviations, numbers, and other formatted text is expressed so that ASR engines can easily use these files as training or testing data. Standardizing the formatting of output is essential for reproducible measurements of ASR accuracy.

split_audio_file

usage: split_audio_file [-h] [--target-dir TARGET_DIR] audio_file transcript

Split an audio file using valid segments from a transcript file. For this
utility, transcript files must contain start/stop times.

positional arguments:
  audio_file            input audio file
  transcript            transcript

optional arguments:
  -h, --help            show this help message and exit
  --target-dir TARGET_DIR
                        Path to target directory

prepare_audio_corpora

usage: prepare_audio_corpora [-h] [--target-dir TARGET_DIR]
                             corpora [corpora ...]

Copy and organize specified corpora into a target directory. Training,
testing, and development sets will be created automatically if not already
defined.

positional arguments:
  corpora               Name of one or more directories in directory this
                        script is run

optional arguments:
  -h, --help            show this help message and exit
  --target-dir TARGET_DIR
                        Path to target directory

This script scrapes a list of directories for paired STM and SPH files. If train, test, and dev folders are present, these labels are used for the output folder. By default, a target directory of 'input-data' will be created. 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.

degrade_audio_file

usage: degrade_audio_file input_file1.wav input_file2.wav

Degrade audio files to 8 kHz format similar to G711 codec

This script reduces audio quality of input audio files so that acoustic models can learn features from telephony with the G711 codec.

extract_excel_spreadsheets

Note that the use of this function requires the separate installation of pandas. This can be done via pip install pandas.

usage: extract_excel_spreadsheets.py [-h] [--input-folder INPUT_FOLDER]
                                     [--output-corpus OUTPUT_CORPUS]

convert a folder of excel spreadsheets to a corpus of text files

optional arguments:
  -h, --help            show this help message and exit
  --input-folder INPUT_FOLDER
                        input folder of excel spreadsheets ending in .xls or
                        .xlsx
  --output-corpus OUTPUT_CORPUS
                        output folder for storing text corpus

align_json

This aligns a gk hypothesis json file with a reference text file for creating forced alignment STM files for training new ASR models. Note that this function requires the installation a few extra packages

python3 -m pip install spacy textacy https://github.com/explosion/spacy-models/releases/download/en_core_web_sm-2.2.0/en_core_web_sm-2.2.0.tar.gz#egg=en_core_web_sm

usage: align_json.py [-h] input_json ref output_filename

align a gk json file against a reference text file

positional arguments:
  input_json       input gk json file
  ref              reference text file
  output_filename  output_filename

optional arguments:
  -h, --help       show this help message and exit

Requirements

• Python >= 3.6.1 with pip

Contributing

Code of Conduct

Please make sure you read and observe our Code of Conduct.

Pull Request process

1. Fork it
2. Create your feature branch (git checkout -b feature/fooBar)
3. Commit your changes (git commit -am 'Add some fooBar')
4. Push to the branch (git push origin feature/fooBar)
5. Create a new Pull Request

NOTE: Commits and pull requests to FINOS repositories will only be accepted from those contributors with an active, executed Individual Contributor License Agreement (ICLA) with FINOS OR who are covered under an existing and active Corporate Contribution License Agreement (CCLA) executed with FINOS. Commits from individuals not covered under an ICLA or CCLA will be flagged and blocked by the FINOS Clabot tool. Please note that some CCLAs require individuals/employees to be explicitly named on the CCLA.

Need an ICLA? Unsure if you are covered under an existing CCLA? Email [e-mail]

Authors

• Matthew Goldey
• Tejas Shastry
• Amy Geojo
• Svyat Vergun
• Ashley Shultz
• Colin Brochtrup

License

The code in this repository is distributed under the Apache License, Version 2.0.

Copyright 2020 GreenKey Technologies