The method of speech recognition based on representation of speech information by 2D time frequency feature vectors stream is considered. These vectors are constructed from the time-frequency information using 2D-wavelet transform of spectrogram image. Classification of vectors is processed by ANN. Sonograms of short time Fourier transform and adaptive Hermite transform form are the input speech information. These representations are compared for tasks of speaker-independent speech recognition and speech-independent speaker recognition.
