An in-service non-intrusive measurement device for characterising speech networks

The work presented in this thesis evaluates echo path modelling using an in-service non-intrusive method. The in-service non-intrusive measurement devices (INMDs) are based on least mean square (LMS) digital adaptive filters (DAFs). The modelling convergence rate (misadjustment) is derived from the optimal Wiener weights, and is used to define the performance criterion and the excitation for the DAFs is conversational speech. One second adaptation is allowed before the DAF's tap weights are interrogated in order to determine the echo path response. Hence all improvements quoted are based on one second adaptation. Speech driven INMDs produced a 'step' change in misadjustment during adaptation in noise free conditions. The unvoiced segments in speech produced a faster convergence compared to the voiced segments. In a noisy environment, however, the low energy unvoiced segments are masked by noise, thus producing divergence. Novel techniques were developed to minimise divergence in a noisy environment. The methods developed include divergence detectors (DDs) and an adaptive step-size algorithm. Implementing DDs revealed that a long running energy based detector produced a better divergence minimisation and showed an improvement of 28 dB, in terms of misadjustment, at echo to noise ratio (e/N) of 0 dB after one second adaptation. Meanwhile, the adaptive step-size algorithm where the instantaneous step-size is derived from the correlation of the input speech, showed a 31.3 dB improvement under the same conditions. The investigation of the LMS convergence properties revealed that, white signals gave the fastest convergence rate. Hence, three new whitening techniques were designed, Fast Fourier Transform Least Mean Square (FFTLMS), parallel DAF measurement device (PDMD) and chaotic LMS. The FFTLMS method employed the reciprocal input spectrum to flatten the power spectral density (PSD) spread and yielded an improvement of 22.2 dB, while the PDMD method makes use of the LMS stochastic behaviour upon convergence to provide similar whitening effect and an improvement of 23.8 dB was achieved. Both methods involve the pre-filtering the input and echo path speech with whitening coefficients generated by the respective methods. The other method proposed in this thesis is the chaotic LMS, where the input speech in chaotically coded before adaptation. The coded speech has a white spectral density and higher overall energy contents. The performance using the coded speech is similar to white noise, and the improvement achieved was 34.3 dB after 125 ms.