A Configurable Accelerator for Keyword Spotting Based on Small-Footprint Temporal Efficient Neural Network

Abstract

Keyword spotting (KWS) plays a crucial role in human–machine interactions involving smart devices. In recent years, temporal convolutional networks (TCNs) have performed outstandingly with less computational complexity, in comparison with classical convolutional neural network (CNN) methods. However, it remains challenging to achieve a trade-off between a small-footprint model and high accuracy for the edge deployment of the KWS system. In this article, we propose a small-footprint model based on a modified temporal efficient neural network (TENet) and a simplified mel-frequency cepstrum coefficient (MFCC) algorithm. With the batch-norm folding and int8 quantization of the network, our model achieves the accuracy of 95.36% on Google Speech Command Dataset (GSCD) with only 18 K parameters and 461 K multiplications. Furthermore, following a hardware/model co-design approach, we propose an optimized dataflow and a configurable hardware architecture for TENet inference. The proposed accelerator implemented on Xilinx zynq 7z020 achieves an energy efficiency of 25.6 GOPS/W and reduces the runtime by 3.1× compared with state-of-the-art work.

A Resource-Efficient Keyword Spotting System Based on a One-Dimensional Binary Convolutional Neural Network

“…MFCCs are widely used to represent the audio features [2][3][4][5][6]9,10,12,14,[20][21][22]. The calculation process for the MFCCs can be described as follows.…”

“…The CNN model presented in [22] was used to perform classification with MFCC-based feature maps. The 1D-CNN model presented in [5] was used to find features to recognize the specific part of a keyword. The long short-term memory (LSTM) model presented in [21] was trained to capture long-term information in an audio stream.…”

“…To reduce the complexity of the KWS system, various techniques have been proposed to optimize the feature extraction and classification process. K. He [5] and Y. S. Chong [6] attempted to reduce the complexity of calculating the MFCCs. In the work of K. He [5], the discrete cosine transformation and windowing procedure were removed.…”

“…K. He [5] and Y. S. Chong [6] attempted to reduce the complexity of calculating the MFCCs. In the work of K. He [5], the discrete cosine transformation and windowing procedure were removed. In the work of Y. S. Chong [6], a rectangular filter was introduced instead of the conventional triangular filter to reduce computational and memory requirements.…”

9.1 µW keyword spotting processor based on optimized MFCC and small‐footprint TENet in 28‐nm CMOS

“…In this letter, a low-power, high-accuracy KWS architecture is proposed. In comparison to previous work [3], we integrate the feature extraction circuit [4] corresponding to the simplified MFCC algorithm as the frontend input for the accelerator. The architecture in this study is fully implemented using a 28 nm process technology.…”