No abstract

Automatic speech recognition (ASR) enables very intuitive human-machine interaction. However, signal degradations due to reverberation or noise reduce the accuracy of audio-based recognition. The introduction of a second signal stream that is not affected by degradations in the audio domain (e.g., a video stream) increases the robustness of ASR against degradations in the original domain. Here, depending on the signal quality of audio and video at each point in time, a dynamic weighting of both streams can optimize the recognition performance. In this work, we introduce a strategy for estimating optimal weights for the audio and video streams in turbo-decodingbased ASR using a discriminative cost function. The results show that turbo decoding with this maximally discriminative dynamic weighting of information yields higher recognition accuracy than turbo-decoding-based recognition with fixed stream weights or optimally dynamically weighted audiovisual decoding using coupled hidden Markov models.

No abstract

Proposed is a new approach for the separation of localised sources in a reverberant environment, using ad hoc distributed microphones. In a first stage, the method uses a fuzzy clustering algorithm to assign microphones to individual sourcedominated 'clusters'. This assignment is used to generate time-frequency masks to form an initial estimate of the source signal at each microphone of the corresponding cluster. Since microphone positions are not typically available in ad hoc arrays, the initial separation is exploited to derive the parameters required for delay-and-sum (DSB) beamforming. This beamforming is subsequently carried out, utilising all the microphones within a source cluster. Following this, a timefrequency post-filtering mask is estimated via the exchange of power spectra of the beamformed signals. This mask is applied to the beamformer outputs to further improve the separation. The approach is tested in three realistically-simulated rooms of different sizes, and evaluated by informal listening tests as well as by instrumental quality metrics (comprising both quality and intelligibility measures). Our experiments demonstrate that the approach can lead to a significant separation of the individual sources, yielding results very close to that obtained by DSB's where oracle knowledge of microphone and source positions is incorporated.

This contribution deals with the in situ detection and localisation of brake squeal in an automobile. As brake squeal is emitted from regions known a priori, i.e., near the wheels, the localisation is treated as a hypothesis testing problem. Distributed microphone arrays, situated under the automobile, are used to capture the directional properties of the sound field generated by a squealing brake. The spatial characteristics of the sampled sound field is then used to formulate the hypothesis tests. However, in contrast to standard hypothesis testing approaches of this kind, the propagation environment is complex and time-varying. Coupled with inaccuracies in the knowledge of the sensor and source positions as well as sensor gain mismatches, modelling the sound field is difficult and standard approaches fail in this case. A previously proposed approach implicitly tried to account for such incomplete system knowledge and was based on ad hoc likelihood formulations. The current paper builds upon this approach and proposes a second approach, based on more solid theoretical foundations, that can systematically account for the model uncertainties. Results from tests in a real setting show that the proposed approach is more consistent than the prior state-of-the-art. In both approaches, the tasks of detection and localisation are decoupled for complexity reasons. The localisation (hypothesis testing) is subject to a prior detection of brake squeal and identification of the squeal frequencies. The approaches used for the detection and identification of squeal frequencies are also presented. The paper, further, briefly addresses some practical issues related to array design and placement.