Principal Investigator
|
Project Title
| Time-Frequency filtering for the removal of overlapping speech |
|---|
Brief Description for General Publications
This study is intended to concentrate on the study of time-frequency based techniques for the removal of overlapping speech. As discussed below, this requires high performance computing. A large amount of work has been carried out in the past forty years on the development of better hearing aids, initially the work was aimed at improving analog circuits for the hearing aids. In recent years, the majority of work has aimed at improving digital signal processing (DSP) techniques for use in hearing aids. Modern hearing aids are programmed at the time of fitting to utilise all of the hearing capabilities of a patient. They feature multiple compensation channels, which allow the gain to change across the spectrum. The combination of both these features utilise's the full capabilities of the modern DSP chips. Digital signal processing techniques provide the possibility of adaptive online processing compared to the fixed frequency band design developed for analogue hearing aids. There is still a great need for the research to be done into developing hearing aids in which competing voices are filtered out and compensation is applied to the "target" voice to overcome the hearing loss of the subject. The motivation for the research undertaken in this project is to develop an algorithm for hearing aid applications in which not only is the hearing loss of the patient addressed but also voice filtering aimed at reproducing the cocktail party affect. A hearing aid with either voice filtering or patient specific amplification will improve the patient's speech perception. Implementation of both voice filtering and patient specific amplification will improve a patient's speech perception further since they compliment each other. The technique under development uses a continuous wavelet transform technique to represent the characteristics of each individual speaker. A speaker is characterised by a time-frequency basis of wavelet filter templates that are used to select for the desired speech characteristics and separate out the unwanted speech signals. In order to develop a practical representation of typical speech it is necessary to correlate a large number of speaker examples in order to determine a minimal sufficient basis for the adaptive identification of speaker characteristics for an on-line algorithm. This research will draw from earlier work undertaken by the computer science laboratory at ANU that recorded and catalogued native Australian english speakers from all over Australia. Each speech sequence will be processed by a wavelet bank chosen in order to capture possible speech frequency components and time localisation characteristics. The resulting data will be correlated across all speakers to provide a global picture of speaker characteristics. This information will be used to provide a guide to the choice of a minimal set of speaker characteristics in wavelet space for the development of on-line algorithms on specialised DSP architectures. |