Advanced Query by Humming SystemQBH Advanced Query by

Advanced Query by Humming System(QBH) 電信所一 王治皓

參考 • Advanced Query by Humming System Using Diffused Hidden Markov Model and Tempo Based Dynamic Programming (Chiao-Wei Lin, Jian-Jiun Ding, and Che-Ming Hu) • IMPROVED ONSET DETECTION ALGORITHM BASED ON FRACTIONAL POWER ENVELOPE MATCH FILTER (Jian-Jiun Ding , Che-Ming Hu , Ta Hsien , and Chi-Jung Tseng ) • YOUTUBE影片

大綱 1. onset detection(起始檢測) 2. pitch estimation(音調提取) 3. melody matching(旋律匹配)

INTRODUCTION 1. 把信號分割成一個一個音符 signal note

ONSET DETECTION 困難處 • 當信號能量變化不大時，難以檢測出onset point • Thrill(顫抖音)和 end tone of the music signal 容易使我們誤判 onset point

PROPOSED METHOD (FOR ONSET DETECTION) 1. 2. 3. 4. 5. Difference of magnitude Short-term energy HFC method Surf method Proposed method 與方法 1較類似

(Step 1)find the envelope amplitude for each time slot. 類似於Difference of magnitude

(Step 2)Take the fractional power of the envelope amplitude. 假設Ak = 0. 12, Ak 1 = 0, Ah = 0. 52, and Ah 1 = 0. 4. = 0. 7 Dk =Ak –Ak 1 =0. 12 Dh = Ah –Ah 1 =0. 12 Bk =0. 1995 Bh =0. 1062

(Step 3)跟match filter做convolution. f[n] = [3, 3, 4, 4, -1, -2, -2, -2]. (Step 4) 如果convolution的值大於thd，我們就可以把第K點當 作onsetpossible 1.

(Step 5) Ck : convolution of envelope and envelope matched filter Max(i) and Min(i) : the i-th peak in Ck.

Hidden Markov Model • 傳統HMM的小麻煩 • 解決方案 : 使用diffusion matrix

10 -second case. Method DP HMM LS [14] [15] [16] Proposed Method MRR 0. 725 0. 577 0. 639 0. 728 0. 750 0. 742 0. 774 20 -second case. Method DP HMM LS [14] [15] [16] Proposed Method MRR 0. 803 0. 743 0. 754 0. 831 0. 941 0. 929 0. 983