Neural network mapping of throat microphone speech
Nasim Radmanesh, Zahir M. Hussain · RMIT Research Repository (RMIT University Library) · 2009
Throat microphone is robust to the surrounding noise and can even pick up whispers; however, speech recorded from throat microphone (TM) is unnatural and metallic, unlike normal microphone (NM) speech. The aim of this paper is to improve the quality of the throat microphone speech via efficient mapping of TM speech spectra into NM speech. The TM speech is typically a low bandwidth signal whereas the NM speech is of wider bandwidth. Thus, to improve throat microphone speech, the missing frequencies should be added to speech spectra. In this paper, algorithm for finding linear predictive (LP) coefficients and cepstral coefficients are mentioned. Also, A multilayered feedforward back-propagation neural network is used to map feature vectors (cepstral coefficients) of the two speech signals.