In this article considered with problem talker-independent recognition 
of phonemes in speech that occurs when creating automatic word recognition 
of discrete and continuous speech based on phonemic-based method. 
To solve this problem the neural network recognition algorithm phonemes. 
The material of speech corpus TIMIT performed experiments aimed 
at evaluating the accuracy of the developed neural network algorithm 
in recognizing English phonemes of continuous speech. 
A comparative analysis of the experimental data with the results obtained 
by using hidden Markov models.