We consider the new technology of speech corpus creating.
It is proposed to use the existing recognition systems 
to build transcripts of speech signal.
Expert compares the audio recording and the automatic transcript and marks incorrectly words.
For further train exclusively correctly recognized speech signal is used.
The proposed segmentation method allows to obtain speech corpus without 
direct human stenography, which greatly reduces the workload 
of experts ang their qualifications to effectively get great speech corpus 
without attracting highly qualified expert linguists.
The experiments confirmed the efficiency of the developed technology.