使用聲調模型輔助之基頻偵測器與國語連續語音聲調辨認

Abstract

在本論文中，我們提出了使用聲調模型來輔助基頻軌跡的估測，並在估測軌跡的同時辨認每個音節的聲調。以往的基頻軌跡估測常發生半頻與倍頻錯誤，於是我們提出使用具有統計性的聲調模型輔助軌跡估測，希望可以利用模型的統計特性，使得半頻與倍頻錯誤的發生減少，在實驗中我們發現，使用模型輔助估測軌跡的方式，確實可以減少上述兩種錯誤。以往在聲調辨認方面，均是先求取基頻軌跡，然後利用估測後的基頻軌跡進行聲調辨認，而在本文中提出一種架構，此架構可以同時估測基頻軌跡與聲調辨認。

In this thesis, a new model-based pitch tracking scheme is proposed. It tracks pitch and recognizes tones simultaneously using a statistical prosody model of Mandarin speech. With the guide of the prosody model, the pitch tracking can be more reliable so as to reduce both half pitch error and double pitch error. Experimental results showed that the gross pitch error (GPE) of pitch detection was reduced from 1.121% to 0.918% by using the proposed pitch estimation scheme. Both half and double pitch error rates were also reduced. Meanwhile, a tone recognition rate of 70% was achieved.