EUROSPEECH 2003 - INTERSPEECH 2003
The present public telephone networks trim o. the lowband (50-300 Hz) and the highband (3400-7000 Hz) components of sounds. As a result, telephone speech is characterized by thin and muffled sounds, and degraded speaker identification. The lowband components are deterministically recoverable, while the missing highband can be recovered statistically. We develop an equalizer to restore the lowband parts. The highband parts are filled in using a linear prediction approach. The highband excitation is generated using a bandpass envelope modulated Gaussian signal and the spectral envelope is generated using a Gaussian Mixture Model. The mean log-spectrum distortion decreases by 0.96 dB, comparing to a previous method using wideband reconstruction with a VQ codebook mapping algorithm. Informal subjective tests show that the reconstructed wideband speech enhances lowband sounds and regenerates realistic highband components.
Bibliographic reference. Qian, Yasheng / Kabal, Peter (2003): "Dual-mode wideband speech recovery from narrowband speech", In EUROSPEECH-2003, 1433-1436.