Abstract [eng] |
Purpose of this project is to analyze Kaldi toolkit possibilities in automatic speech recognition researches. The most widely audio collection was used, named SKAIC30 - 30 speakers voice recordings with Lithuanian numbers from 0 to 9, which was expanded to 100 speakers voice recordings with 5dB background noise. The final work presents a comparison with the results of the HTK software package using 30 voice recorders. For further research, a 100 speakers voice recordings with 5dB background noise was used, in order to check which method of the following: monophone, triphone, LDA+MLLT, LDA+MLLT+SAT, SGMM, DNN provides the most accurate results for automatic speech recognition system. |