Table 4 Word success rate (in percent) for a simulated STNO neural network with \(N=2000\) nodes after filtering the inputs with the cochlear filter combined with the reservoir.

From: Role of non-linear data processing on speech recognition task in the framework of reservoir computing

SNR (dB)SubwayBabbleCarExhibitionAVG
clean88.81 (+23.88)87.21 (+24.81)89.02 (+29.02)91.51 (+25.87)89.14 (+25.90)
2067.54 (+23.51)69.77 (+20.16)69.02 (+26.67)56.76 (+22.01)65.77 (+23.09)
1570.15 (+27.61)63.57 (+16.96)68.63 (+23.14)59.46 (+28.57)65.45 (+24.07)
1055.97 (+18.66)53.49 (+14.73)56.86 (+21.17)53.28 (+21.62)54.90 (+19.05)
AVG70.62 (+23.42)68.51 (+19.17)70.88 (+25.00)65.25 (+24.52)68.82 (+23.02)
  1. These results are for a reservoir trained with all noise levels and types and then tested for each part of the test set with the different noise levels and types. Each entry gives the overall word recognition rate and in parentheses, the gain achieved by adding the reservoir to the preprocessing filter. The first column gives the noise level for different parts of the test set and the top row gives the noise type. The bottom row gives the results over all levels of noise for each noise type and the right column gives the results for each noise level over all types of noise. The bottom right entry is the average result over the whole test set.