Joint Dictionary Learning-Based Non-Negative Matrix Factorization for Voice Conversion to Improve Speech...

Area of Science:

• Machine Learning
• Speech Processing
• Bioacoustics

Background:

• Surgical removal of articulators can lead to distorted speech, impacting patient communication.
• Existing voice conversion (VC) methods face challenges with limited training data and require rapid conversion for practical use.

Purpose of the Study:

• To develop an effective machine learning-based voice conversion (VC) technique for improving speech intelligibility in surgical patients.
• To address the limitations of small training datasets and the need for efficient conversion in post-operative communication.

Main Methods:

• A novel joint dictionary learning based non-negative matrix factorization (JD-NMF) algorithm is proposed.
• The JD-NMF method is designed for efficient and effective VC with limited training data.

Main Results:

• The JD-NMF method significantly improves short-time objective intelligibility (STOI) scores compared to original speech.
• Experimental results show JD-NMF is more efficient and effective than conventional exemplar-based NMF VC methods.

Conclusions:

• The proposed JD-NMF method demonstrates superior performance in enhancing speech intelligibility for oral surgery patients.
• The joint training criterion for NMF-based VC is validated, confirming the effectiveness of JD-NMF.