In real-time signal processing there is a constraint to finish processing of an audio signal before the next audio segment is received. This makes it important to have signal processing algorithms with low computational complexity while still maintaining high quality results. This thesis presents methods for audio signal processing used in real-time systems. The publications presented cover areas of noise reduction, network echo cancellation, noise dosimeter measurements and voice analysis.

A method for speech enhancement is presented with low amounts of speech distortion. The audio signal is split into several subbands, covering different frequency regions. For each subband, the noise level is estimated. A signal gain is calculated by comparing the total signal level with the noise level for each subband. The method presented here, improves performance compared to previously similar methods. Improvement is especially found in multi-speaker and noise-only scenarios.

When communicating on a telephone line, network echo is introduced by hybrids in the network. In cases where multiple echo sources exist, the time range for echoes can be quite long. In devices with limited storage, it is difficult to get good echo cancellation in such cases. This thesis presents a method for network echo cancellation suited for use in a device with a larger external memory.

Exposure to high noise levels will have negative health effects and methods for measuring noise level exposure is important. Included in this thesis is a study that remove the influence of own voice in noise dose measurements.

For certain medical conditions it is beneficial with daily voice exercises. Methods for grading voice in four different exercises is presented, based on pitch and loudness. Evaluation is done in real-time on test medical device.

This paper presents an asynchronous resampling implementation on a low-power fixed-point DSP, which uses around 47% less computational resources compared to the solution provided by the DSP manufacturer, without compromising audio quality.

In high-quality conferencing systems, it is desired to perform noise reduction with as limited speech distortion as possible. Previous work, based on time varying amplification controlled by signal-to-noise ratio estimation in different frequency subbands, has shown promising results in this regard but can suffer from problems in situations with intense continuous speech. Further, the amount of noise reduction cannot exceed a certain level in order to avoid artifacts. This paper establishes the problems and proposes several improvements. The improved algorithm is evaluated with several different noise characteristics, and the results show that the algorithm provides even less speech distortion, better performance in a multi-speaker environment and improved noise suppression when speech is absent compared with previous work.

Long delays and sparseness characterize impulse responses in telecommunication networks and a vast number of solutions for network echo cancellation have been proposed over the years. In this paper, an approach for detecting dispersive regions of a sparse impulse response and a proportionate normalized least mean square (PNLMS)-based selective updating approach are combined with an adaptive double-talk detector to form a complete solution for echo cancellation. The proposed solution has low computational complexity and is targeted for systems equipped with external memory.

This paper presents a personal voice analyzer and trainer that allow the user to perform four daily exercises to improve the voice capacity. The system grades how well the user is performing the exercises by analyzing the duration, the intensity and the pitch of the user’s voice.

Noise dosimeters are valuable tools in assessing the individual noise dose in the workplace. At non-industrial work places with a high degree of communication, such measurements would include the wearer’s own voice which would be registered as noise. This may not always be desirable. The purpose of this investigation was to study the effect of the wearers own voice in noise dosimeter measurements, and especially the difference between children and adults as test subjects. The study took place at a day-care center and sixteen children and thirteen adult female preschool teachers participated. The participants wore a digital recorder during the day, which recorded the sound signal and vibrations originating from an accelerometer attached to the neck of the test subjects, for distinguishing of whether the subject was speaking or not. Thus, average A-weighted noise levels with and without the influence of the subjects own voice could be obtained. The Leq for the measurements with and without the own voice was 84.6 dBA and 72.2 dBA for the children, respectively, and 79.3 dBA and 70.0 dBA for adults. Student’s t-test showed a significant (p<0.01) difference of 12.4 dBA for children and 9.3 dBA for adults when comparing measurements including and excluding the own voice and also that the difference was significantly larger for children. Thus, the study conclude that the influence from the own voice implied an augmentation of the Leq value and that there is a significant difference between children and adults in how large this augmentation is.