Improvement of Esophageal Speech using LPC and LF Model

Esophageal speech is a restoration of speech communication in laryngectomized patient. Due to irregular pharyngoesophageal (PE) segment vibration and aerodynamic limitation, the esophageal phonation provides higher volatility of the fundamental frequency (f ₀) compared with laryngeal phonation. It is difficult to determine an accurate f ₀ in esophageal speech. This paper focuses on algorithms for f ₀ modification. Linear predictive coding (LPC) and autocorrelation function are used to calculate the f ₀ They are well performed in the normal case. However, the determination results of f ₀ in esophageal speech are highly unstable without any modification of a conventional LPC technique. By proposing a smoothing technique, an accurate f ₀ and pitch period in esophageal speech can be determined and used in LF model. Experimental results from 18 subjects suggest that average f ₀ of esophageal phonation is lower than laryngeal phonation. The speech synthesized using this proposing technique produced better sound quality than un-processed esophageal speech.