Power efficient implementation of MVA-SI on speech controlled IoT systems

Speech processing is implemented based on application specific accuracy requirements and regardless of the implementation scheme the ultimate goal remains the same, to enhance speech processing engines up to an extent where natural language can be interpreted accurately in case of word or phoneme recognition and context prediction. ASR training and adaption is always application specific and in the considered application the ASR engine is trained for a power efficient IoT application. The nodes communicate via 802.15.4 6LoWPAN which makes wireless sound transmission very lightweight and power efficient. MVA-SI is a customized version of i-vector speaker identification algorithm which identifies the speaker with system wake up call. In this paper, the logic and implementation of MVA-SI are discussed in detail along with elaboration of prototype system architecture.