Training a Speech-to-Text Model for Dutch on the Corpus Gesproken Nederlands
Willem Röpke · VUBIR (Vrije Universiteit Brussel) · 2019
Speech-to-text, also known as Speech Recognition, is a technology that is able to recognize and transcribe spoken language into text.In subsequent steps, this transcription can be used to complete a multitude of tasks, such as providing automatic subtitles or parsing voice commands.In recent years, Speech-to-Text models have dramatically improved thanks partially to advances in Deep Learning methods.Starting from the open-source project DeepSpeech, we train speech-to-text models for Dutch, using the Corpus Gesproken Nederlands (CGN).First, we contribute a pre-processing pipeline for this dataset, to make it suitable for the task at hand, obtaining a ready-to-use speech-to-text dataset for Dutch.Second, we investigate the performance of Dutch and Flemish models trained from scratch, establishing a baseline for the CGN dataset for this task.Finally, we investigate the issue of transferring speech-totext models between related languages.In this case, we analyse how a pre-trained English model can be transferred and fine-tuned for Dutch.