Removing several kinds of unexpected burst noises before conventional noise reduction process is very useful in many noise environmental speech processing areas, because it can not only reduce the processing time and system memory, but improve the system performance and efficiency.
In this thesis, we suggest a new algorithm that can remove different types of sporadic noises occurring in daily-life office environment. Our algorithm basically utilizes both the human speech properties and sporadic noises to differentiate sporadic noises from the input signal.
As a research material, 21 kinds of office environment sporadic noises and 17 isolated words are collected. The total numbers of noise and word data are 918 and 357, respectively. The isolated words include 6 words consisting of voiced and unvoiced sounds, 6 words consisting of voiced sounds only, and 5 words having a consonant at the end of utterances.
By applying our new algorithm, 402 sporadic noises are successfully removed by the simple rule utilizing the basic properties of human speech, while 463,23,10 noises are eliminated by utilizing spectral variance, cumulant, and pitch information, respectively. As a result, 898 office environment sporadic noises among 918 are removed without any damage to the speech utterances, so the noise canceling success rate is 97.82%.
Test results also show that noises with simple shape pattern and short duration are relatively easy to remove. That means noises with rather complicated shape pattern or long duration are somewhat hard to cancel successfully. From our experimental result, we can also observe that in case of relatively long duration noises, even the whole cancellation is not successfully achieved, considerable portion of each noise is eliminated in many cases. Thus, we can say that our proposed algorithm is more reliable than the success rate mentioned above.
As a conclusion, we strongly believe that our algorithm can be utilized to improve the perfo...