Implementing template matching logic in a standard flash memory cell

Abstract

Current research into classification methods is almost exclusively software based, resulting in systems that perform well but are invariably slow when faced with large databases. The goal is therefore to create a hardware classification system that is much faster. In this paper, we introduce the concept of template matching logic and propose the use of a standard flash memory cell array to perform bit by bit template matching. The proposed system is based on a novel architecture that is unique and separate from existing architectures that make use of flash memory cell arrays. Verification is achieved by speech recognition simulations on the TIMIT database. Simulations of the system show results of 94.5 % recognition accuracy on clean words and 88.0 % recognition accuracy on test words with a signal-to-noise ratio of 5 dB. The results compare favorably to similar isolated word recognition tasks performed with software based methods.

Memristor pattern recogniser: isolated speech word recognition

“…To apply the speech data to the proposed memristor cell array, we encode the analogue value x to binary of B bits using a simple digitisation process [5] where X is the digitised matrix, k is the index denoting the coding bit ( k = 1, 2, 3, …, B ), and thresh is a vector containing the thresholds for each bit level The memristor model used for the simulation was proposed by Kvatinsky et al [6]. The model is driven with a pulsed input of ± 3.5 V using the following fitting parameters: R ON = 100 Ω, R OFF = 2 × 10 5 Ω, frequency = 20 kHz, source amp = 0.003 A, D = 3 nm, P coeff = 2, initial state = 0.5, j = 1.5, a on = 2.3 nm, a off = 1.2 nm, i on = −1 µA, i off = 1 µ A, X c = 107 × 10 −11 m, k on = −8 × 10 −13 m/s, k off = 8 × 10 −13 m/s, , x on = 0, α off = 3, α on = 3, and window_type = Kvatinsky.…”

“…Switches S R will close during the READ phase when required inputs such as V T generated through the test input voltage are applied to the cell. To apply the speech data to the proposed memristor cell array, we encode the analogue value x to binary of B bits using a simple digitisation process [5]…”

“…These voltage levels, ±3.5 V, are used to set high-and low-resistance states of memristor and are represented as ± V TR , the training voltages. The TIMIT database was used for word data filtering from sentences sa1 and sa2, adding white Gaussian noise to replicate a more realistic scenario, and feature extraction using standard signal processing techniques to convert the input waveform into a Bark-scale spectrogram [5]. Specifically, the waveform was divided into sections of length 25 ms, windowed with a Hamming window, with an overlap of 10 ms between adjoining sections.…”