german_transliterate

german_transliterate is a Python module to clean and transliterate (i.e. normalize) German text including abbreviations, numbers, timestamps etc. It can be used to clean messy text (e.g. map peculiar Unicode encodings to ASCII) or replace common abbreviations in text in combination with various text mining tasks.

However, it is particularly useful for Text-To-Speech (TTS) preprocessing (both in training and inference) and has features to support phonemic encoding of the results (e.g. with espeak-ng) afterwards as next step in the processing pipeline.

Is has been successfully applied to preprocessing with Mozilla TTS in combination with espeak-ng phonemes as input data to both training and inference pipeline.

License and Attribution

This work is licensed under the Creative Commons Attribution 4.0 International License. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/ or send a letter to Creative Commons, PO Box 1866, Mountain View, CA 94042, USA.

To provide attribution or cite this work please use the following text snippet:

german_transliterate, Copyright 2020 by repodiac, see https://github.com/repodiac for updates and further information

Version History

• 0.1.3 - some bugfixes in various ops: weekday, month, amount_money and acronyms, also some minor stuff fixed here and there (update highly recommended)
• 0.1.2 - removed the following ops from the list of default ops, since (as mentioned in the documentation below) they are highly error-prone (many false-positives). You can still use them via explicitly adding them to the transliterate_ops=[...] list. The ops removed are:
  • month
  • weekday
  • math_symbol
• 0.1.1 - added command-line interface for default usage (no phoneme encoding and experimental stuff left out)
• release 0.1 - initial release of the software, still a lot of ToDos and some more experimental features (see documentation); also exception handling could be improved

Installation/Setup

It has currently only one external dependency, num2words. All dependencies are to be found in requirements.txt and included in setup.py as well, at the moment.

Installation is easy using pip and built-in git package installation based on setup.py:

• pip install git+https://github.com/repodiac/german_transliterate

Setup:

• It should install and behave (import german_transliterate.core) to your current Python environment as any other pip package (in case, create a virtual environment with virtualenv or conda before).

Documentation

Example Usage

In Python code or as library:

from german_transliterate.core import GermanTransliterate

text = 'Um 13:15h kaufte Hr. Meier (Mitarbeiter der Firma ABC) 1.000 Luftballons für 250€.'
print('ORIGINAL:', text, '\n')

ops = {'acronym_phoneme', 'accent_peculiarity', 'amount_money', 'date', 'timestamp',
        'weekday', 'month', 'time_of_day', 'ordinal', 'special', 'math_symbol', 'spoken_symbol'}

# use these setting for PHONEMIC ENCODINGS as input (e.g. with TTS)
print('TRANSLITERATION with phonemic encodings:',
      GermanTransliterate(replace={';': ',', ':': ' '}, sep_abbreviation=' -- ').transliterate(text), '\n')

# use none or your own for other purposes than phonemic encoding and do not use 'spoken_symbol' or 'acronym_phoneme'
print('TRANSLITERATION (default):',
      GermanTransliterate(transliterate_ops=list(ops-{'spoken_symbol', 'acronym_phoneme'})).transliterate(text), '\n')

NEW From command-line (in the shell):

python core.py '1, 2, 3 - alles ist dabei'

Input Parameters

There is currently only one method to be used: transliterate('Das ist der Text.')

It has the following input parameters:

• transliterate_ops list of keywords, see below for details
• replace dict of "original: replacement" string tuples to be used as additional plain and simple "on-the-fly" replacements with the text, e.g replace={'-' : ' '} replaces all dashes with whitespace; leave empty for normal use and use {';': ',', ':': ' '} with phonemic encodings
• sep_abbreviation a special separator used for transliteration of abbreviations; this is mostly only useful with phonemic encoding of a text as a next step in a TTS pipeline; leave empty for normal use and use ' -- ' with phonemic encodings
• make_lowercase if True, text is made lowercase (leave empty by default) NOTE: most of the transliterate operations do only work with make_lowercase=True - this is due to the various dictionaries operating with lowercase only. Please use make_lowercase=False only when transliterate_ops aren't overly used, otherwise most of them do not work!

The parameters used for the config parameter transliterate_ops are as follows:

• acronym_phoneme transliterates abbreviations like ABC into a phonemic version ah beh zee
• accent_peculiarity removes peculiar Unicode encodings and maps them to compatible ASCII-like versions (cleaning op)
• amount_money transliterates currency and money symbols like $, €, EUR etc.
• date transliterates dates, e.g. 12.10.2019
• timestamp transliterates timestamps, e.g. 13h:15m:45s
• weekday (experimental), transliterates abbreviations for weekdays, for instance Mo -- currently this is rather error-prone (many false-positives)
• month (experimental), transliterates abbreviations for months, e.g. Jan or Dez -- currently this is rather error-prone (many false-positives)
• time_of_day transliterates time of day, e.g. 13:15h
• ordinal transliterates ordinal numbers, e.g. 2. into zweite (tries to find a tradeoff for correct case suffix, i.e. zweiten or zweitem)
• special transliterates edge cases and special terms, e.g. 8/10 into acht von zehn
• math_symbol (experimental), transliterates a small selection of math symbols, e.g. plus, minus etc. (also here applies: can have a lot of false-positives, so use with care)
• spoken_symbol allows to transliterate brackets or citation marks into spoken language, e.g. '( text )' into -- in klammern -- text -- (if sep_abbreviation is set to ' -- '), mainly useful for TTS tasks

Performance

The current state is mainly based on using manual mappings and regular expressions for substitution and expansion of strings (words or terms). Therefore, current performance should be good enough to be used with online inference or "realtime" usage in a text processing pipeline. As further modules or ops are added over time, there might be also rather slow methods doing heavy computations and thus suited mainly for training or offline processing.

Issues and Comments

Please open issues on github for bugs or feature requests. You can also reach out to me via email.