Implementing a Statistical Parametric Speech Synthesis System for a Patient with Laryngeal Cancer

Area of Science:

• Speech synthesis
• Medical acoustics
• Computational linguistics

Background:

• Total laryngectomy significantly impacts patient quality of life due to voice loss.
• Laryngeal cancer patients experience dysphonia, affecting communication and psychosocial well-being.
• Developing assistive voice technologies is crucial for post-laryngectomy rehabilitation.

Purpose of the Study:

• To create a statistical parametric speech synthesis system for a laryngeal cancer patient before total laryngectomy.
• To evaluate the quality of synthetic speech generated from the patient's pre-surgery recordings.
• To assess the potential of this technology to improve the quality of life for patients awaiting laryngectomy.

Main Methods:

• A statistical parametric speech synthesis model was trained using the Merlin repository and a Polish language corpus.
• Speech samples from a laryngeal cancer patient, recorded before surgery, were used for training.
• Auditory-perceptual MUSHRA listening tests were conducted with 25 experts to rate synthetic voice quality.

Main Results:

• The synthetic voice generated from the patient's recordings achieved a high MUSHRA score (69.4/100).
• This patient-derived synthetic voice outperformed a synthetic voice trained on a professional voice-over talent's recordings (63.63/100).
• The patient's pre-surgery voice exhibited dysphonia, confirmed by RBH scale and Acoustic Voice Quality Index (AVQI) analysis.

Conclusions:

• A statistical parametric speech synthesizer can be effectively created for patients awaiting total laryngectomy.
• This technology offers a promising solution to restore voice quality and enhance the mental well-being of patients.
• The developed system demonstrates the feasibility of generating high-quality synthetic speech tailored to individual patient needs.