語音增強使用感知激勵頻譜振幅之貝氏估計器

Speech Enhancement Using Perceptually Motivated Bayesian Estimators of the Magnitude Spectrum

高士喆(Shih-Je Kao)

指導教授 ： 簡福榮

國立臺北科技大學/電資學院/電機工程系研究所/碩士(2014年)

https://doi.org/10.6841/NTUT.2014.00925

摘要

語音信號會受到背景雜訊的影響而導致整體的語音品質降低，而語音增強系統的主要目的為降低背景雜訊對語音訊號的影響。藉由使用雜訊預估演算法搭配語音增強演算法來重建乾淨的語音訊號，使增強後的語音訊號有比較低的語音失真。
本篇論文中使用了兩種感知激勵頻譜振幅預估語音增強演算法與三種雜訊預估演算法，其中語音增強演算法有加權歐機里德(Weighted Euclidean)和加權雙曲餘弦(Weighted Hyperbolic Cosine)，雜訊預估演算法包含最小統計法(MS)、最小控制遞迴平均法(MCRA)與改善式最小控制遞迴平均法(IMCRA)。實驗結果顯示，使用本文的感知激勵頻譜振幅預估方法與溫尼濾波器和最小均方誤差(MMSE)相比，感知激勵頻譜振幅預估方法在SSNRI、SDI和NRF可以獲得較好的改善。其中又以使用加權雙曲餘弦(Weighted Cosh)增強演算法搭配最小控制遞迴平均(MCRA)雜訊預估演算獲得最明顯的增強效果。

關鍵字

語音增強 ； 雜訊預估 ； 感知激勵語音增強 ； 貝氏估計器

並列摘要

Speech signals are tend to decrease the speech quality when corrupted by background noises. The main purpose of speech enhancement systems is to reduce the background noise from a noisy speech signal by using both noise estimation algorithm and speech enhancement algorithm, such that produce an enhanced speech signal that has a relatively low speech distortion.
In this thesis, we develop two speech enhancement algorithms based on perceptually motivated estimators of the magnitude spectrum together with three noise estimation algorithms. Speech enhancement algorithms include Weighted Euclidean (WE) and Weighted Hyperbolic Cosine (WCOSH). Noise estimation algorithms include Minimum Statistics (MS), Minima Controlled Recursive Averaging (MCRA), and Improved Minima Controlled Recursive Averaging (IMCRA). The experimental results show that compared with the Wiener filter and Minimum Mean-Square Error (MMSE), the perceptually motivated estimators provide better SSNRI, SDI and NRF scores. Among all, the Weighed Cosh method incorporated with MCRA achieves the most significant enhancement performance.

並列關鍵字

Speech Enhancement ； Noise Estimation ； Perceptually-Motivated Speech Enhancement ； Bayesian Estimators

參考文獻

[1] M. S. Choi and H. G. Kang, “A two-channel noise estimator for speech enhancement in highly non-stationary environment,” IEEE Transactions on Audio, Speech, and Language Processing, vol. 19, no. 4, May 2011, pp. 1-11.

[2] S. Boll, “Suppression of acoustic noise in speech using spectral subtraction,” IEEE Transactions on Audio, Speech, and Signal Processing, vol. 27, no. 2, Apr. 1979, pp. 113-120.

[5] Y. Ephraim and D. Malah, “Speech enhancement using a minimum mean-square error short-time spectral amplitude estimator,” IEEE Transactions on Acoustics Speech and Signal Processing, vol. ASSP-32, no. 6, Dec. 1984, pp. 1109–1121.

[6] Y. Ephraim and D. Malah, “Speech enhancement using a minimum mean-square error log-spectral amplitude estimator,” IEEE Transactions on Acoustics, Speech and Signal Processing, vol. ASSP-33, no. 4, Apr. 1985, pp. 443–445.

[7] P. C. Loizou, “Speech enhancement based on perceptually motivated Bayesian estimators of the speech magnitude spectrum,” IEEE Transactions on Speech and Audio Processing, vol. 13, Sept. 2005, pp. 857-869.