Preparation
To further focus the challenge and match demand to supply, an extensive market consultation was carried out in collaboration with Addestino.
Market consultation
The actual market consultation, organised by PIP together with the Flemish Parliament and Addestino, took place in the Flemish Parliament on 24 November 2017.
More than 35 participants, from - both domestic and foreign - companies and knowledge institutions, were present.
By means of an overview of 'use cases' or challenges (downloadable below), they showed what the technology has to offer in the short and long term to convert speech into text.
Addestino recorded the results of this workshop in a report. The report (in Dutch) and a summary (in English) are available below.
Test day
As a follow-up to the market consultation of 24 November 2017, the Flemish Parliament and PIP organised a test day in the Flemish Parliament on 4 May 2018.
The aim of this test day was to validate the current technological possibilities (and their quality) of automatic transcription from Dutch-language speech to text and to further adapt them to the needs of the reporting services of the Flemish Parliament (= Proof Of Concept [POC]).
From the applications submitted for participation in the test day, three participants were selected, as was communicated in advance in the invitation. On the test day itself, they received three audio fragments (of about half an hour each) unknown to them and without accompanying speakers and reports, which they had to convert into textual reports using their speech-to-text model. At the end of the test day, the texts generated by the model were delivered. These texts were assessed by three reporters of the Flemish Parliament on the basis of the following criteria: (1) the time needed for the reporters to convert each piece of text into a 'finished report', (2) the type of errors or passages to be corrected (whether or not related to a specific speaker or grammatical construction, whether or not 'easily' semiautomatically correctable, etc.), (3) the percentage of accuracy of the speaker recognition, and (4) the similarity with the original report, according to the estimation of the editor who made this original report.
The conclusions of the test day were incorporated into a report available below.