Speech-Recognition Techniques for Robust Acoustic Signature Identification

Military target vehicles of different models have unique acoustic signatures. These signature can be used for detection and for identification. One approach is to analyze the signature in the spectral domain, using the Fourier transform, and comparing new sound samples with known templates. Problems arise when signatures vary or are corrupted by noise. The problem of signature recognition has much in common with speech recognition, an active field for several decades. We propose to apply traditional and novel speech-recognition techniques to the identification of acoustic signatures. Preliminary tests in this proposal, applied to military vehicle signature data, show very high accuracy rates. In Phase I we will first develop "traditional" speech recognizers, which use non-Fourier features that have proven more robust, such as cepstral and linear-predictive-coding (LPC) coefficients. As back ends, we will investigate both hidden Markov models (HMMs) and neural networks. The algorithms will be modified to fit the problem domain in terms of frequency ranges and time scales. We will also apply more recent algorithms: the auditory image model (AIM) mimics the human auditory system; learning vector quantization (LVQ) is a novel neurally inspired, Bayesian optimal algorithm. The systems will be compared to the more traditional Fourier-based approach and evaluated for noise robustness. In Phase II, we will develop a noise-robust, real-time system to be evaluated in the field.