Low-complexity feature extraction unit for “Wake-on-Feature” speech processing

Event-Driven Continuous-Time Feature Extraction for Ultra Low-Power Audio Keyword Spotting

“…Between 2 successive filtering stages, an additional max-pooling layer is inserted. At the end, a fully connected layer, with a size matching the number of target keywords, is set after the last filtering step [1]. This kind of CNN classifiers is compatible with low-power embedded implementation [12].…”

“…out-of-band attenuation, relaxes the complexity constraints. For speech recognition, it has been demonstrated in [1] that 8 filters logarithmically distributed between 75Hz and 7kHz can extract sufficient features to recognize 10 keywords with 90% accuracy. Since the human voice does not have a significant spectral content below 800Hz (only information on voice intonation and accent are present) [8], we propose to first restrict the frequency range of interest from 75Hz-7kHz to 840Hz-6.25kHz to reduce hardware complexity.…”

“…In such systems, the problem appears to be the implementation of the feature extraction unit with low hardware complexity level, ultra-low energy consumption as well as a reduced number of extracted features to simplify the classification mechanism [1]. In literature, several methods have been proposed to design the feature extractor.…”

“…The event-driven operation builds a proportionality between dynamic power consumption and the activity of the input signal, what makes the system suitable for the recognition of keywords present in audio signals with long silence periods. The principal hardware-oriented parameters (such as number of bands, ADC resolution and energy quantizer resolution), are optimally chosen to scale down the system complexity while maintaining a recognition accuracy of 90%, as proposed in [1] for a discete-time digital system. Since in CT-DSP systems, only finite impulse response (FIR) filters with a limited number of taps are feasible [6], this paper proposes to study the implementation of a filter bank composed of 8 CT FIR filters, with a maximum of 16 delay taps each.…”

On the Distribution of Clique-Based Neural Networks for Edge AI

“…Pre-processing data in the sensor aims at decreasing the use-time of the communication interface by determining the relevance of the input data in regard to the targeted application. If the acquired data is relevant, a "wake-up" signal can be generated to enable the transmission; otherwise, the communication interface is in sleep mode, as illustrated in [3]. Moreover, when the interface is activated, only pre-processed data can be transmitted for analysis instead of raw data, thereby reducing the number of transmitted bits further.…”