Speech reconstruction using a deep partially supervised neural network.

Area of Science:

• Speech processing
• Biomedical engineering
• Machine learning

Background:

• Dysphonia treatment often relies on statistical speech reconstruction methods.
• Gaussian mixture models and restricted Boltzmann machines have shown promise.
• Deep neural networks (DNNs) are limited by small patient datasets.

Purpose of the Study:

• To develop a novel DNN structure for speech reconstruction in dysphonia.
• To enable effective training with limited voice data from individual patients.
• To improve upon current state-of-the-art speech reconstruction techniques.

Main Methods:

• Proposed a novel deep neural network (DNN) architecture.
• Implemented a partially supervised training approach.
• Utilized spectral features from small datasets for training.

Main Results:

• Achieved very good performance in speech reconstruction.
• Demonstrated superior results compared to existing methods.
• Successfully trained DNNs on limited patient data.

Conclusions:

• The novel DNN structure is effective for speech reconstruction in dysphonia.
• Partially supervised training overcomes data limitations in DNNs.
• This approach offers a promising solution for voice-loss patients.