Classification of imagined spoken word-pairs using convolutional neural networks

EEG-Based Silent Speech Interface and its Challenges: A Survey

“…In contrast, TL assumes their domain is different, or the future labelling task may differ. Some studies reported that TL did not decrease the accuracy [36], [37]. However, as the accuracy was still poor (35-60%), it needs further observation.…”

“…Similar to Arizona State University's dataset, the studies on Coretto's dataset also achieved the highest accuracy by implementing a deep learning model (30% for vowel classification [35] and 62.37% for word classification [36]). It was validated by a general CV, as shown in Table V.…”

“…The signals can be aligned to the onset of imagined speech production and then averaged to focus between the preceding and following onset time [59]. Previous studies have examined it and found that the potential in the left hemisphere arises onetwo seconds before speech production [36] and appears a moment before cued speech is produced [60]. One study reported that speech-related ERP reached the peak at 350ms after the cue [12], and the highest significant level (from the paired t-test) existed between 400-600ms after the cue in the frontal area [17].…”

“…Common Spatial Patterns (CSP) and Principal Component Analysis (PCA) are other extraction methods. Filter bank CSP (FBCSP) [61] was the state-of-theart for feature selection [35] or classification [36], [37], [62].…”

“…Riemannian manifold RVM CV [27] channel cross-covariance (CCV) CNN+LSTM+DAE CV [28] channel cross-correlation matrix LSTM CV [29] DWT DNN CV [30] Bag of Features (BoF) RNN LOSO-CV [30] Visual (Spain): "arriba" ("up"), "abajo" ("down"), "izquierda" ("left"), "derecha" ("right"), "seleccionar" ("select") 4-25 [31] 4-25 DWT RF CV [32] 0-64 Statistics, RWE [33] 40-50 Bag of Features (BoF) NB + TL CV [34] Audiovisual (Spain) "/a/", "/e/", "/i/", "/o/", "/u/", "arriba" ("up"), "abajo" ("down"), "izquierda" ("left"), "adelante" ("forward"), "atrás" ("backward") 2-40 RWE RF CV [35] CNN CV [36] CNN layer, FBCSP [37] CNN + TL LOSO-CV [38] word embedding + Siamese encoder kNN CV [39] Visual "/a/", "/e/", "/i/", "/o/", "/t/" 0.5-220 phase per band (Hilbert transform) SVM CV [40] Audio "/a/", "/e/", "/i/", "/o/", "/u/", "yes", "no", "left", "right". 0.1-70 RMS, zero-crossing rate, moving window average, kurtosis, and power spectral entropy RNN CV [41] Audio "Hi Bixby", "Call Mom", "Open Camera", "What's the weather".…”