Modelling the nonstationarity of speech in the maximum negentropy beamformer

State-of-the-art automatic speech recognition (ASR) systems can achieve very low word error rates (WERs) of below 5% on data recorded with headsets. However, in many situations such as ASR at meetings or in the car, far field microphones on the table, walls or devices such as laptops are preferable to microphones that have to be worn close to the user’s mouths. Unfortunately, the distance between speakers and microphones introduces significant noise and reverberation, and as a consequence the WERs of current ASR systems on this data tend to be unacceptably high (30-50% upwards). The use of a microphone array, i.e. several microphones, can alleviate the problem somewhat by performing spatial filtering: beamforming techniques combine the sensors’ output in a way that focuses the processing on a particular direction. Assuming that the signal of interest comes from a different direction than the noise, this can improve the signal quality and reduce the WER by filtering out sounds coming from non-relevant directions. Historically, array processing techniques developed from research on non-speech data, e.g. in the fields of sonar and radar, and as a consequence most techniques were not created to specifically address beamforming in the context of ASR. While this generality can be seen as an advantage in theory, it also means that these methods ignore characteristics which could be used to improve the process in a way that benefits ASR. An example of beamforming adapted to speech processing is the recently proposed maximum negentropy beamformer (MNB), which exploits the statistical characteristics of speech as follows. “Clean” headset speech differs from noisy or reverberant speech in its statistical distribution, which is much less Gaussian in the clean case. Since negentropy is a measure of non-Gaussianity, choosing beamformer weights that maximise the negentropy of the output leads to speech that is closer to clean speech in its distribution, and this in turn has been shown to lead to improved WERs [Kumatani et al., 2009]. In this thesis several refinements of the MNB algorithm are proposed and evaluated. Firstly, a number of modifications to the original MNB configuration are proposed based on theoretical or practical concerns. These changes concern the probability density function (pdf) used to model speech, the estimation of the pdf parameters, and the method of calculating the negentropy. Secondly, a further step is taken to reflect the characteristics of speech by introducing time-varying pdf parameters. The original MNB uses fixed estimates per utterance, which do not account for the nonstationarity of speech. Several time-dependent variance estimates are therefore proposed, beginning with a simple moving average window and including the HMM-MNB, which derives the variance estimate from a set of auxiliary hidden Markov models. All beamformer algorithms presented in this thesis are evaluated through far-field ASR experiments on the Multi-Channel Wall Street Journal Audio-Visual Corpus, a database of utterances captured with real far-field sensors, in a realistic acoustic environment, and spoken by real speakers. While the proposed methods do not lead to an improvement in ASR performance, a more efficient MNB algorithm is developed, and it is shown that comparable results can be achieved with significantly less data than all frames of the utterance, a result which is of particular relevance for real-time implementations.