返回Google圖書搜尋

Echo cancellation and noise reduction

Bogomir HorvatZdravko KačičAndrej Miksić

其他書名	final report [for Siemens]
出版	Research and Studies Center University of Maribor, 1996
URL	http://books.google.com.hk/books?id=ftbvOAAACAAJ&hl=&source=gbs_api

註釋Speech recognition in adverse environments is a very important issue for real-life automatic man-machine communications. It was the aim of this project to further evaluate on the topic of echo cancellation for the purpose of automatic telephone dialog systems, and to add the spectral substraction noise reduction capatibiltiy. In the echo cancellation part, the work was therefore an upgrade to the previous project, which is in detail described in the Final Report, Project Echo Cancellation, October 1995. The new features added now are the stability and robustness to low-energy echo, a novel duble-talk detection method, and adjustment of all parameters for optimal recognition performance as tested on large amount of speech material and on real echo signals recorded with analog and digital handsets. Much research was done to confirm the reduced-subband approach, where the number of frequency bands is reduced from the number of frequency bins after the fft to 24 frequency bands. Experiments on speech endpoint detection were also conducted. Furthermore, experiments simulating the stopping of the system prompt in the double-talk conditions were carried out. Another task was to include noise reduction module to the frontend processing. Some work on that topic was already done at Siemens, and now the module was integrated in a way that both noise reduction and echo cancellation can be combined. For both echo cancellation and noise reduction, the integer version of the software were written and the results with the integer version do not differ much from the floating point implementation. All the experiment results are given in section Experiments and results.