以過濾法從語音信號之線性預測殘留來擷取聲門關閉瞬間

A Filtering Method for Extracting Glottal Closure Instants from Linear Prediction Residual of Speech Signal

胡懷祖(Hwai-Tsu Hu)

《宜蘭技術學報》 3期 (1999/12) Pp. 69-85

https://doi.org/10.6179/nit.1999.03.18

摘要

本文提出一種以過濾方法來告示潛藏於低通過濾後之殘留信號中的重大變動訊息。由於聲門關閉瞬間是以負波峰的型態出現於處理過的殘留信號，我們便試圖增強和搜尋各式語言殘留信號的負波峰，納入實驗驗證的案例則以一些過去引發問題的單音節為主，這包括有/u/、/m/、/n/、/z/、/a^w/、/Λ/與/η/。而這種過濾法於白雜訊場合的穩健性亦是我們的探討重點，其效能是以一個低音階的母音/u/於10 dB信噪比的情況作為示範，若與其它三種方法比較，我們所提出的方法最能在聲門關閉瞬間呈現明顯對比。此外，集合24句語料庫所得到的實驗結果亦指出此法在一般信噪比環境的工作情形令人滿意。而聲門關閉瞬間的準確判斷不僅有助於擷取語言信號之聲學特徵，更便於以音高同步的方式來處理語音。

關鍵字

聲門關閉瞬間 ； 線性預測殘留 ； 過濾

並列摘要

A filtering method is proposed to signify the epochal feature residing in the lowpass filtered residual. While the glottal closure instant (GCI) manifests itself as a negative peak in the processed residual, our method aims at the enhancement and retrieval of these peaks for a large variety of speech signals. The voiced sounds considered in our experiments include /u/, /m/, /n/, /z/, /a^w/, /Λ/, and /η/, all of them were reported to be problematic. The robustness in the presence of white Gaussian noise is also under our investigation. Its performance is demonstrated by trying out a low-pitched vowel /u/ with SNR=10dB. Compared with three other methods, the proposed method produces an evident contrast at the GCI's. Furthermore, the results based on a database consisting of 24 sentences indicates that this method works well in a moderate SNR environment. The accurate determination of GCI's not only helps with the extraction of acoustic features of speech signals, but facilitates the application of a pitch synchronous approach to speech processing.

並列關鍵字

Glottal closure instant ； Linear prediction residual ； Filtering