Enhancing Real-Time Automatic Speech Recognition for Flemish Speech (CRISP)

This project aims to enhance an existing AI model developed by the ESAT - PSI lab at KU Leuven for transcribing Flemish speech, including dialects and non-native speakers, to function in quasi real-time (1) and be deployable on edge devices (2). Current state-of-the-art models, like Whisper from OpenAI, struggle with variations such as regional dialects and accents, as well as speech from non-native speakers. Our ASR model, however, excels in these areas and has a smaller footprint, making it suitable for edge devices such as smartphones or MR devices.
To reach the abovementioned goals, PSI will further streamline the data-processing pipeline of the developed ASR model. Flanders Make will evaluate the model in the context of the Human-Machine Flexible Production use case, investigate edge deployment, and help the industry partners to deploy it into their applications. Other potential uses include automatic logging systems for repair workers, teaching assembly procedures through verbal explanations, and helping non-native Dutch speakers learn Dutch.

Downloads

nelfproject (NeLF Project) 

Contacts

Abdellatif Bey-Temsamani (Flanders Make) - abdellatif.bey-temsamani@flandersmake.be

Researchers

Hugo Van hamme (KULeuven) - hugo.vanhamme@kuleuven.be
Ewoud Verhelst (Flanders Make)

Research Groups

KULeuven ESAT - PSI

Flanders Make 

NeLF