This script converts a textual input containing spelled out numbers into one where numeric values are written as numbers. It is suitable only for DUTCH. Spelled out numbers in the input are expected to be deconstructed in a way that is typically used in automatic speech recognition, e.g., twee en twintig, honderd veertien. The output will not conform to the rules from Onze Taal, but this could be implemented as an option at a later time.
Input can be plaintext, .ctm file format, or JSON strings in a format that is compatible with Kaldi's output. This script could for example be included in a full postprocessor pipeline for a Kaldi-based live recognition system.
When using .ctm or JSON input, the results should have the correct timestamps, durations, and posteriors. For JSON input, multiple hypotheses are supported.
Usage:
echo 'Op vier juli twee duizend achttien was het mooi weer in postcodegebied vijf en zestig drie en veertig' | ./words2num_nl.py
Results in:
Op 4 juli 2018 was het mooi weer in postcodegebied 6543
Some special cases are recognized, such as pairs of two-digit numbers which can be combined into one four-digit number as these often refer to either calender years or zip-codes in Dutch. Also the word 'komma' between numbers is interpreted as a decimal symbol.
Other cases, such as monetary values, time of day information, etc have not been implemented (yet).