Digital Signal Processing Reference
In-Depth Information
Illustration 295:
Prognosis speech recognition
Here, the successful prognosis of the created neuronal network is showing. All 50 voice samples have been
recognized without mistake. Still, the tolerance of the neuronal network isn´t high, it is strongly speaker-
dependent. The reason particularly lies in the physical background. Professional applications use the
entire two- dimensional “bed- of- nails” pattern of each word. Here in contrast only the one- dimensional
spectrum is analysed. 4096 spectral values per word are being reduced to 8 frequency parameters by
signal preprocessing.
Now, the correct network prognosis can either be tested with a preliminary microphone
or by reading an according multi file. If the result is unsatisfying, the constellation for the
frequency parameter module has to be changed and the following steps have to be
repeated.
As a definite characteristic of the successful use of a neuronal network, the
distinguishability of the training data regarding the recognizable target values has already
been mentioned. In this case, the Illustration of the spectra in a frequency- time- landscape
already shows this distinguishability and also wether the training material is suitable for
proceeded process.
