Automatic classification of the acoustical situation using amplitude-modulation spectrograms
Jürgen Tchorz, Birger Kollmeier · The Journal of the Acoustical Society of America · 1999
A fast and reliable classification of the acoustical situation is an important prerequisite for modern speech processing schemes. Noise suppression algorithms in digital hearing aids, for example, are strongly dependent on a proper noise level estimation. The classification algorithm which is presented is based on so-called amplitude-modulation spectrograms (AMS). It is motivated by neurophysiological findings in the auditory cortex in mammals. Its basic idea is that both spectral and temporal information of the signal is used to attain a separation between ‘‘acoustical objects’’ within the signal. Spectral information is gained by splitting the input signal into different frequency bands, temporal information is gained by analyzing amplitude modulations in each frequency band. Thus for each analysis frame of the input signal, a two-dimensional AMS pattern is created. Signals with similar spectral shape but different temporal properties are clearly distinct in the AMS representation. A neural network is trained on a large number of AMS patterns which were generated from speech and noise samples. After training, an automatic classification of speech and noise from ‘‘unknown’’ sources is performed with high accuracy. Furthermore, an estimate of the present signal-to-noise ratio is supplied for mixed input signals.