Runtime Reconfigurable Keyword Spotting for Low-power Microcontrollers

The human-machine interaction through audio interface, has gained a lot of attention in recent years. Many application, like google assistant, Amazon Alexa, Siri, are nowadays everywhere in the world and work only through vocal interaction. The motivation of the abrupt diffusion of these systems is the ease of accessing the functionalities of such devices through the natural voice. The trend for the future seems to be a pervasive diffusion of these systems. Typically, a small local device is responsible of collecting the audios and transmits them to a remote server, where a complex system, able to translate the audio into actions, is implemented. An enabling factor, for a correct implementation of this paradigm, has been represented by KeywordSpotting (KWS). KWS are applications that mount on the local device and act as a filter, forwarding to the cloud only the audios that are actually valuable. The principle through which these systems work is based on the spotting of an activation word. When a KWS recognizes the activation word, it activates the remote application and begin the communication of the request of the user. The practise of implementing KWS is an enabling factor for drastically reducing the power consumption. KWS applications are implemented locally over edge devices like MCUs. They are composed of two stages, a \acrfull{fe} stage that is responsible of performing spectral analysis over the input and of a classification stage that is usually implemented through a deep learning model. Implementing both these stages inside a limit resources device like MCUs, is a challenging operation for the correct deployment of a KWS application. For this reason, many optimizations have been proposed in the literature. In recent years, a lot of attention has been put in dynamic re-configuration of deep learning models, over their implementation in edge devices. These works are very promising. The flexibility that is guaranteed by dynamic systems is much more suitable with respect to static techniques, at least on edge devices. In particular, dynamic systems are able to provide, when needed, high accuracy and when necessary, low power consumption. However, these solutions cannot be applied to KWS ecosystem. This, due to the fact that both the feature extraction stage and the deep learning model contribute equally to the consumption of the resources of a device. Optimizing only the deep learning model may lead to only partial benefits for the entire system. In this contest, the thesis want to propose one of the first example of a dynamic KWS system, with the emphasis of optimizing both the stages that constitute a KWS application. Far from proposing an optimal solution, the thesis aims at introducing the concept of full dynamic optimization in the KWS ecosystem. The hope is that in the future, more literature could benefit from this thesis. The thesis proposes a dynamic power-driven system: FDO, that is able to provide a configuration of SotA accuracy (91.87\%) and a configuration of low power consumption, that is able to spare half the power consumption with respect to the high accuracy configuration. The thesis also proposes an input-driven system that is build upon FDO: AdaptFDO. The system is able to provide a pareto curve of possible run time configurations, each of them being a trade-off between latency and accuracy. The power savings, with respect to the high accuracy configuration, range from x1.14 to x1.41, while providing a small loss in term of accuracy.