A Special Speaker-dependent Speech Enhancement Algortihm

With the development of information technique, voice communication becomes basic method for communication. However, in practical life, voice communication is seriously affected by various kinds of noise in environments. Thus, both the performance of speech processing system and the quality of voice communication are degraded dramatically due to this noise influence. Therefore, many researchers focus on efficient speech enhancement algorithm. Meanwhile, electronic products are usually utilized by few fixed persons nowadays. Based on this feature, we try to combine special speaker’s feature and speech enhancement algorithm to improve performance.In this master thesis, we propose a special speaker-dependent speech enhancement algorithm. With speech absence and presence models, optimal modified minimum-mean-square-error log-spectral amplitude(OMLSA) speech estimation combined with improved minima-controlled recursive averaging(IMCRA) noise estimation performs better than other conventional algorithms. So we modify IMCRA-OMLSA algorithm based on special speaker’s feature.In speaker-dependent speech enhancement algorithm, speaker’s Gaussian mixture model(GMM) is trained as special speaker’s feature. According to the influence of speech presence probability on different estimation processes and weighting parameters, speech enhancement is modified by speaker’s GMM feature from three aspects: priori signal-to-noise rate estimation, noise estimation and speech estimation.The performance of proposed speaker-dependent speech enhancement algorithm is evaluated by six objective tests under various noise types and signal-noise-ratio. From the experimental results, compared to the conventional approaches, the proposed special speaker-dependent algorithm performs better at the aspect of both noise reduction and speech distortion.