Security Instrument using Talker Identification and Microphone Arrays in Variable Noisy Environments

In paper, we propose a robust security instrument that can detect the location and identity of talker in variable noisy environment. This instrument calculates the value of the signal-to-noise (SNR) using a discrete-valued SNR estimator. This SNR value is used to adapt the performance of speech/nonspeech classifier to the surrounding noisy environment. If the detected signal is speech then a novel multi-engine talker identification (TT) will determine the identity of the talker, else an audio classification system will determine the audio type of the detected signal (e.g. widows breaking, and wind sounds). Multi-engine TI utilizes the SNR value to select the TI engine, from a set of six engines (five different SNR environments and a "clean" environment), with the SNR training condition that best matches the surrounding environment; greater TI accuracy should be achieved when training and test environments are similar. The performance of this instrument is evaluated using twelve test environments, with the SNR ranging from -10 dB to clean environment (SNR > 50 dB). The proposed instrument achieves an average classification accuracy of 92% over an SNR range of 10 dB to clean environments; an enhancement of 38% over the instrument trained in a clean environment.