Figure 2
Spoken digit recognition for filtered inputs. (a) Spoken digit cross-validated test recognition rates as a function of the number of data subsets \(N\) used for training (total size of the training set \(5\times 10\times N\)) of the filtered input (without neural network) corresponding to four different methods: cochleagram, MFCC filter, Spectro HP and linear spectrogram \((\alpha =1)\). (b) Spoken digit recognition as a function of non-linear coefficient for spectrogram methods (Inset: Word success rate for large non-linear coefficient values from 1000 to 1004). Here, 9 data subsets (90% of the database) are used for training our reservoir computing model and the remaining subset (10% of the database) is used to perform the recognition task. The shaded region corresponds to the uncertainty of the recognition rate, here the standard deviation).
