Almost all the papers that report work on speech enhancement add known amount of noise and then use their proposed algorithm to enhance the speech or reduce the noise level. Here again, mostly noise is assumed to be white Gaussian. However, if one records speech on the road or a market place, there is no guarantee that the noise is Gaussian. For an enhancement algorithm to be really useful in life, it must improve the perceived quality of speech that was originally noisy due to some environmental conditions and NOT due to explicit addition of noise by the researcher. I would like to know studies, where people have tested their algorithms on such real noisy speech.
Although work on the Serbian language, you can use these reference works that treat nongaussian noise. The results are excellent.
The base of real interference is available at the URL http://www.utdallas.edu/ ~ Loizou / speech.
http://2009.telfor.rs/files/radovi/09_47.pdf
http://2008.telfor.rs/files/radovi/08_28.pdf
http://infoteh.etf.unssa.rs.ba/zbornik/2009/radovi/B-III/B-III-10.doc
Best regards and success.
Prof. dr Zoran Velickovic
The work of Ji Ming et al might be relevant. The method reconstructs speech using a corpus of clean speech, so it needs no knowledge of the noise. If you have access to IEEE Explore, see http://dx.doi.org/10.1109/TASL.2010.2064312 for the paper 'A Corpus-Based Approach to Speech Enhancement from Nonstationary Noise'
- Badges
- Science topic
- Similar topics
- Correction