Speech signal separation and synthesis based on a scene analysis audi tori and modeling

摘要

Systems and methods for producing clear speech from a speech signal, which consists of a mixture of noise and speech, are provided. The clear language can be produced from synthetic speech parameters. The synthetic speech parameters are based on the speech signal components and a language model with the use of acoustic and speech derived production principles. The model can be a sources filter structure of the speech signal use. One or more spectral analyzes of the speech signal are carried out, in order to generate spectral representations. The characteristic data are derived based on a spectral representation. The properties which, according to a language model of the target language are grouped together and separated from the characteristic data. The synthetic speech parameters, including the spectral envelope, voting position data and voting classification data based on properties are produced, which correspond to the target language.