Neural networks in separation of voice signal from audio record

U ovom radu dan je pregled trenutnog stanja područja izdvajanja govora i pregled najuspješnijih strategija za izdvajanje govora. Poseban naglasak stavljen je na duboke neuronske mreže. Odabrana je BLSTM-RNN arhitektura neuronske mreže i CURRENNT kao programski paket za rad s tom mrežom. Uspješnost izdvajanja govora je ispitana na skupu podataka prvog zadatka drugog CHiME natjecanja (engl. CHiME 2nd challenge Task 1) i postignuto je apsolutno poboljšanje točnosti od 25% za prepoznavanje govora sa smetnjama koristeći sustav za prepoznavanje čistog govora s malim opsegom rječnika.An overview of recent developments in the field of speech extraction is given, including a survey of most successful strategies. A special emphasis is put on approaches based on deep neural networks. The BLSTM-RNN neural network architecture is chosen along with the CURRENNT software package for working with the network. Success of speech extraction is evaluated on Task 1 of the CHiME 2nd challenge, and an absolute improvement in word accuracy of 25% is achieved on a noisy speech recognition task using a small-dictionary ASR system specialised for clean speech