Digital Enhancement of Speech Perception in Noisy Environments (Aachener Beiträge zu digitalen Nachrichtensystemen)

In many everyday situations, people try to listen to speech from a loudspeaker, but the speech perception is affected by background noise. The location of a listener is called the near-end. Some applications in which this problem arises are mobile phones, public address systems, car entertainment systems, hearing aids, or head phones. Depending on the application, the speech may either be a recording, or, in case of communication systems, it may come live from a person speaking at the so-called far-end. Although it is usually not possible to cancel the near-end noise source, the speech perception can be enhanced nevertheless by adaptively pre-processing the loudspeaker signal. To find out in which way the speech should be pre-processed, it is beneficial to know the near-end noise characteristics and take them into account. The knowledge can be gained from a near-end microphone that is located close to the listener. This pre-processing technique is called Near-End Listening Enhancement (NELE). In this work several NELE concepts are developed and studied with the goal of enhancing the speech perception. Afterwards, a computationally very efficient and real-time capable NELE algorithm is developed, which can be implemented on signal processors and integrated into low-power devices.
When deploying these enhancement concepts, additional problems may arise depending on the application. Two of them are studied in detail:
1. If the near-end loudspeaker and microphone are placed close to each other, which might occur at public address systems for instance, the speech from the loudspeaker is coupled into the near-end microphone as crosstalk and disturbs the estimation of the noise characteristics. An analysis of this arrangement shows that the functionality of the NELE algorithm is strongly affected. As a solution, it is proposed to use a new noise estimation technique which provides significantly more accurate noise estimates than state-of-the-art approaches by taking the adaptively predicted crosstalk into account.
2. Especially in the case of mobile communication, it is possible that a noise source is present also at the far-end. Consequently, the speech signal from the far-end is disturbed by noise. Typical mobile phones attenuate the far-end noise by employing techniques of noise reduction (NR) before transmitting the signal to the near-end. At the receiving near-end device, the denoised speech signal would then be processed by NELE to improve the speech perception, based only on the near-end noise. It is shown that NR and NELE counteract each other which results in unpleasant speech with strong artifacts. General solutions are proposed to perform NR and NELE under joint control or even jointly at the near-end or the far-end or within the network.