Automatic Speech Recognition in Realistic Conditions

U ovom je radu predstavljena kratka povijest razvoja sustava za prepoznavanje govora. Opisane su osnovne metode koje se koriste u sustavu za prepoznavanje govora i njihova ograničenja: skriveni Markovljevi modeli, dinamičko savijanje vremena, neuronske mreže, end-to-end automatsko prepoznavanje govora. U praktičnom dijelu rada testirana su dva novija end-to-end ASR modela, koja se temelje na dubokim neuronskim mrežama, a to su wav2vec 2.0 i komercijalni Amazonov ASR model. Testiranja su provedena na uzorcima „čistog“ i „onečišćennog“, tj. realističnog govora.This paper presents a brief history of the development of speech recognition systems. The basic methods used in the speech recognition system and their limitations are described: hidden Markov models, dynamic time warping, neural networks, end-to-end automatic speech recognition. In the practical part of the paper, two new end-to-end ASR models were tested, which are based on deep neural networks, namely wav2vec 2.0 and the commercial Amazon ASR model. Tests were conducted on samples of "pure" and "contaminated", ie realistic speech