Master Thesis Speech Target Extraction Based on Directives

Targeted speech understanding in a multi-source auditory environment is crucial for the latest hearing acoustics applications such as ear buds, hearing aids, smart watches, in-car acoustics, etc. In the latest hearing devices, AI-based systems such as recurrent neural networks, LSTMs, beamformer networks are deployed to perform noise cancellation, speaker enhancement, etc. With the introduction of directives, target speaker identification is possible without the need for clean target samples. In your Master thesis, you will investigate target speech extraction with different background noises and develop a target extraction method based on directives. You will evaluate speech extraction methods for hearing aids. You will perform data augmentation on toy datasets and implement pre-processing stages of audio. Furthermore, you will develop neural networks to extract target speech based on directives. In addition, you will extend the developed method to multi-source speech targets. Last but not least, you will gain experience and collaborate in a cross-functional team spanning algorithms, deployment and hardware.