Detection and estimation of acoustical signals using hidden Markov models

The recognition of acoustical signals and the enhancement of speech both play an important role in many communication and signal processing systems. With the increasing portability of these systems, solutions which provide reliable operation in various acoustical environments are of particular interest.

In the present thesis, hidden Markov models (HMM) are shown to be a powerful tool for dealing with the areas mentioned above, approaching them as statistical detection and estimation problems, and putting them on a solid mathematical foundation. Based on this, methods and applications of HMM-based signal detection and estimation for block processing systems are presented.

In the area of acoustical signal recognition, automatic speech recognition systems are the most common application. Since the performance of such systems can decrease significantly in noisy environments, robustness against background noise plays an important role. For single word recognition, this thesis gives a contribution in the form of a robust algorithm for signal preprocessing. It includes an endpoint detection scheme for the determination of word boundaries which allows reliable operation also in noisy environments by using a background noise estimation scheme.

As an application example, an automatic command recognition system for operating speech-controlled devices is described. In addition to background noise disturbances, this application also addresses another problem: Since the command recognition system should always be active, it has to reliably detect and extract the relevant command words out of irrelevant speech data (for example conversations in the room). For this 'word-spotting' problem a solution based on word duration models is presented. The speech recognition system is followed by a validation stage which uses a suitable probability normalization to distinguish the command words from irrelevant signals.

Moreover, this thesis shows that the use of hidden Markov models is not limited to the modeling of speech signals. Using an alarm signal recognition scheme for the profoundly deaf as an example, the HMM-based modeling and recognition of environmental sounds is considered. It is shown that compared to other schemes the use of HMMs gives the best recognition rates.

In the area of speech enhancement, the focus of this thesis lies on single-channel methods for reducing background noise in speech signals. Special emphasis is placed on the HMM-based MMSE estimator, where the statistical properties of the speech and noise signal are described by a speech-HMM and a noise-HMM. These two models can be combined into a composite-HMM which serves as the model for the noisy speech signal and can be used to determine the optimal filter for every noisy speech segment.

It is shown that this scheme can be further improved if additional a priori knowledge of the speech signal is employed. In voiced speech segments, an improvement of the noise reduction can be achieved by using the pitch period which allows the disturbing frequency components between the harmonics of the speech signal to be sufficiently removed. Moreover, it is shown that speech segments with low energy have to be processed separately because the HMM-based scheme does not give a reliable estimation for these segments.

The improvement in the HMM-based scheme obtained through the extensions mentioned is demonstrated using a paired comparison test for speech quality judgement.