LSTM based voice conversion for laryngectomees

This paper describes a voice conversion system designed withthe aim of improving the intelligibility and pleasantness of oe-sophageal voices. Two different systems have been built, oneto transform the spectral magnitude and another one for thefundamental frequency, both based on DNNs. Ahocoder hasbeen used to extract the spectral information (mel cepstral co-efficients) and a specific pitch extractor has been developed tocalculate the fundamental frequency of the oesophageal voices.The cepstral coefficients are converted by means of an LSTMnetwork. The conversion of the intonation curve is implementedthrough two different LSTM networks, one dedicated to thevoiced unvoiced detection and another one for the predictionof F0 from the converted cepstral coefficients. The experi-ments described here involve conversion from one oesophagealspeaker to a specific healthy voice. The intelligibility of thesignals has been measured with a Kaldi based ASR system. Apreference test has been implemented to evaluate the subjectivepreference of the obtained converted voices comparing themwith the original oesophageal voice. The results show that spec-tral conversion improves ASR while restoring the intonation ispreferred by human listenersThis work has been partially funded by the Spanish Ministryof Economy and Competitiveness with FEDER support (RE-STORE project, TEC2015-67163-C2-1-R), the Basque Govern-ment (BerbaOla project, KK-2018/00014) and from the Euro-pean Unions H2020 research and innovation programme un-der the Marie Curie European Training Network ENRICH(675324)