So-Young Jeong

Dual channel noise reduction method using phase difference-based spectral amplitude estimation

We present a dual-channel noise reduction method for small mobile devices. Our method incorporates phase difference between channels into the conventional MMSE spectral amplitude estimator. It is possible to suppress unwanted directional noise signals, whose incident direction is different from that of the target speech signal. Experimental results show that the proposed method outperformed conventional dual-channel noise reduction methods in SNR and PESQ evaluations in a noisy environment.

Adaptive noise power spectrum estimation for compact dual channel speech enhancement

We present a novel noise power estimation method based on smoothed spectral minima tracking and subtractive blocking matrix for dual-channel speech enhancement. By combining spectral characteristics of the noisy mixture signals with spatial null beam-forming, noise over- and under-estimation problem can be substantially mitigated. The proposed noise estimation tested on the real-life nonstationary noise conditions outperformed other single-channel noise estimation based on minima tracking and dual-channel phase error based enhancement algorithm.

Dominant speech enhancement based on SNR-adaptive soft mask filtering

In this paper, we present a SNR-adaptive soft mask filter for multi-channel noisy speech enhancement. Incorporating frame-by-frame spectral magnitude ratios into the time-frequency(T-F) mask filter framework, the adaptive filter can be designed robust to changing environments. Experimental results show that the proposed adaptive mask filter can effectively suppress non-stationary noise components even in a closely-spaced microphone pair. Moreover, the soft mask compressed with sigmoidal nonlinearity can reduce musical noises so that improved PESQ values are obtained.

Semi-blind disjoint non-negative matrix factorization for extracting target source from single channel noisy mixture

We present a semi-blind non-negative matrix factorization(NMF) approach to suppress interference noise signals from a single channel mixture signal. By enforcing a disjointness constraint into the NMF error criterion under the semi-blind denoising framework, it is possible to decompose the mixture spectrogram into target and noise components by minimizing overlaps in the time-frequency domain. Experimental results show that the proposed semi-blind disjoint NMF algorithm can significantly suppress non-stationary noise components in the noisy mixture.

Fast multimedia contents retrieval by partially spoken query

We present novel fast multi-pass decoding strategies for recognizing large named-entities on a low-resource embedded device and thus retrieving MP3 music using spoken query, which contains partial segments of whole music titles and artists. After acoustic-phonetic decoding in the first stage processing, we incorporate word boundary information with phonetic confusion matrix into next stage partial word matching. Then, we rescore candidate phone lists using more complex context-dependent acoustic model, whose outputs are the retrieved songs. We tested our retrieval system to the task of retrieving 1000 songs on a commercial MP3 player and could achieve about 15.5% relative improvements in response time over conventional frame-based multi-pass decoding method without sacrificing recognition rates.