Design of an Approximate Radix-2 FFT Butterfly Unit for LSTM-Speech signal-based Parkinson’s Disease Classifier
摘要
Parkinson’s disease (PD) can be symptomatically detected in its early stage by vocal impairments. A vocal feature-based PD classifier using the long short-term memory-recurrent neural network (LSTM-RNN) model is accurate, reliable, and suitable for early diagnosis. The classification accuracy primarily depends on the feature extraction method. Speech signals are fed framewise to a 256-point radix-2 discrete-in-time fast Fourier transform (R2-DIT-FFT), then to a 26-coefficient mel-frequency cepstral coefficient (MFCC) unit and finally to a discrete cosine transform (DCT) to get 12 features/frame. The majority of the computational complexity is due to the FFT unit. Therefore, an approximate arithmetic-based FFT design can provide higher hardware efficiency without losing the required classification accuracy. The range and accuracy analyses are conducted with different data formats. For this, a MATLAB LSTM-RNN model is trained and validated with the Italian Parkinson’s voice and speech dataset (in.wav format). Thus, an approximate 12-bit customized floating-point (CFP) representation is chosen, and it provides a classification accuracy of 85.34% and an F1 score of 86.61%. Later, the Radix-2 butterfly unit (R2BU) is implemented using 2 multipliers and 3 adders in the proposed data format. This 12-bit CFP multiplier requires 343.93