以巴克频谱失真为基础之语音音调周期估测.docVIP

以巴克頻譜失真為基礎之語音音調週期估測 國科會專題研究計畫成果報告 計畫編號：NSC 90-2213-E-324-025 執行期限：90年8月1日至91年7月31日 主持人：朝陽科技大學資訊工程系 助理教授王德譽 E-mail: dywang@mail.cyut.edu.tw 共同主持人： 計畫參與人員： 一、中文摘要 音高估測在語音信號理中是一個當重要的問題，目前研究不論是時域或頻域之估測，基本上都是藉由比較預估波形與原始波形的關聯性或信號雜訊比來決定音調週期。對於完美清音，簡單的波形比對即可達到正確之音高估測，然而實際語音是時變信號，對於清/濁音變換、共振峰混淆等，若音高估測不當，將使合成之語音品質嚴重下降。故本論文以巴克頻譜失真，比較原始語音及預估諧波頻譜，求得最佳之音高週期。模擬結果發現以巴克頻譜失真為基礎之音高估測，配合正弦語音模型，可有效合成高品質之語音。 關鍵詞：音高估測、巴克頻譜失真，正弦語音模型。 Abstract Pitch detection is an important issue in a variety of speech processing applications. Many pitch detection algorithms (PDAs), both in the time and frequency domains, have been proposed for the voiced/unvoiced detection and pitch abstraction. During highly voiced stationary sections of speech, the pitch period is easily observed using PDAs based on peak detection in the time domain, such as auto-correlation function, zero crossing rate and average magnitude difference function (AMDF). In the frequency domain, the PDAs utilize the harmonic structure of the speech spectrum or the spectral auto-correlation property. All of the proposed algorithms have their limitations, and no presently available PDAs can be expected to give perfectly satisfactory results across a wide range of speakers, applications, and operating environments. In this paper, we provide a pitch detection algorithm based on Bark Spectral Distortion (BSD), in which several known features of the perceptual processing of speech sounds by the human ear are emulated. The experimental results show that the proposed method can provide a more accurate than that of sinusoidal speech model, and the reconstructed speech sounds more naturalness. Keywords: Pitch detection, Bark Spectral Distortion, sinusoidal speech model. 二、緣由與目的 音高估測(pitch detection)在語音信號處理中是一個相當重要的問題，尤其在語音編碼器(Vocoders)[1]中經常假設清音為完美週期性信號以減少編碼位元。對於此問題，已有很多的研究提出並由Hess整理於[2]，大體上可分成時域音調估測及頻域音高估測，時域估測往往利用波形位移前後的關聯性(correlation)找到音高週期。運用的關聯函數如自相關函數(auto-correlation)[3]，或平均振幅差函數(Average Magnitude Difference Function, AMDF)[3]；頻域估測則是計