文本不特定之自動音素分段演算法

Text-Independent Automatic Phone Segmentation Algorithm

本論文提出一個文本不特定循序式偵測音素邊界的演算法，在沒有提供任何已知資訊情況下，建構一個自動音素分段系統。其方法是沿著時間一次只尋找一個候選音素邊界點，找到之後即作確認，經確認後才算是偵測到的音素邊界點而進行音素分段。尋找候選音素邊界點時，採用改變音框長度之離散小波轉換，將小波係數換算成一組頻帶能量，從能量排序之頻帶尋找可能的音素存在範圍，接著以貝式資訊修正準則與正規化頻譜變異函式針對候選音素邊界點進行確認。實驗的語音資料是TIMIT語料庫，實驗結果以F-值與R-值表示其音素分段之正確性，在20ms容忍度下，640個實驗語句得到的平均F-值為72%，R-值為75%。

This paper proposes a text-independent sequential phone boundary detection algorithm. Without any previous knowledge, an automatic phone segmentation system is constructed. The method is to sequentially search for a candidate phone boundary and follow by a verification process. The phone segmentation is accomplished when the phone boundaries are verified. The discrete wavelet transform is applied to variable-length frames. The wavelet parameters are converted to a set of band energies. The energy-sorted bands are then used to search for the candidate phone boundaries. The Bayesian information criterion corrected (BICC) and normalized spectral variation function (SVF) are applied for verifying the phone boundaries. To evaluate this proposed algorithm, the experiment was conducted on TIMIT corpus. The performance of phone segmentation was measured in F-value anf R-value. In the condition of 20-ms tolerance, the average F-value of 640 test utterances is 72% and R-value is 75%.