14-2 ???用???? ?哼?選歌」簡?

qHΰ۪I w uۿqv²

by iP, 2012
AO_HUgG

odzOuqvαҽdҡC@ӨAYnqA²檺覡OgѺqrTӧ_AoǤrT]tqWBqBqmWBMWٵAYaAoǸTäjhOqTAӬOΨӴyzqTTAҥHS٬ meta dataAӺqThO{СA]tD۫ߩMMA䤤̥DnTNOD۫ߡCpGڭ̯ϥΥD۫ߨӷjM֡AoNOҿתue˯v]content-based music information retrieval^CطjMSiHjG

]Apiuۿqv]query by singing/hummingA² QBSH^AX~ӴNܦ@ӫܼsDCgN QBSH izAöi@BЬΡC

Yni QBSHADny{pUG

  1. ϥΪ̪ۿJiBzAHͭVqC
  2. ϥέVqAPƮwqiAX̱񪺤QqC
bĤ@ӨBJANϥΪ̪۸নVqAoӹL{٬lܡ]pitch tracking^CAѡAbܩΰۺqɭԡAڭ̳q`ʿn_ʡA~ಣͶgʪiΡ]SOO^A]n_WߴN٬Wv]fundamental frequency^AqqnAڭ̧ƱWvHɶܪVq]٬Vq^AھڦVqAڭ̤~MƮwqiAX̬ۦqC

Ъ`NA𭵪iγq`SWߩʡA]]㦳WvCAiHլݬݡANAbVWAéCt׻uCvAAiHoıAbouvɡAVOSʪAnOѦYMŮ𪺫tyʩҲ͡AbouvɡAV}liWߩʾ_ʡAe{b~iΤ]NFWߩʡAШUCϨҡC

unn򥻶g]fundamental period^AWvNO˼ơCѤWzϧΥiHݨAܩ㪺WߩʡA]ڭ̥iH[kӧX򥻶gCѩ򥻶gHɶܡA]ڭ̳q`N@sꪺnء]frame^AMAXC@ӭت򥻶gAҦpG

bWϤAڭ̩ҳBzneOuMؤjǸTtvAWvO 16 KHz]]NOCnIO 16000^Cڭ̥X@ӭءA׬O 512 I]ɶ׬O 512/16000 = 0.032 sec = 32 msec^AMϥ[kAboӭؤD3ӧ㪺򥻶gA}l 83 IA 485 IA]򥻶gɶ׬O (485-83)/3/16000 = 0.008375 secAӹWvhO 1/0.008375 = 119.40 HzANCjN 119 Ӱ򥻶gC

ѩڭ̤HչnCPıAäOMnWvAӬOMnWߪƭȦA]ڭ̥iHϥΥbt]semitone^ӪܭApUG $$ pitch = 69 + 12 log_2 \left(\frac{freq}{440} \right) $$ 䤤 $freq$ OH Hz 쪺WvȡA $pitch$ hOH semitone 쪺ȡCϥγoӤANiHڭ̪^^AҦp $freq=440$ ɡAҹ쪺O $pitch=69$AoNO^ La C]^խvq`HeӶiխADnNOeWvNO 440 HzAάO 69 semitoneC^

ϥ[kӧ쭵äAOYnϥιqӦ۰ʧ쭵ANݭn@ǧ޳NFIo̦ܦhkiHΨӧ쭵A̪ı@ؤkA٬۬ơ]audo-correlation functionA² ACF^AzO@ӭؤжi業ΤnA̫X@ ACF uAA즹uĤG̤jȪmAYiXC

]ڭ̨ϥ $s(i)$ ӥNؤ i ӰTȡA ACF iHܦpUG $$ acf(\tau)=\sum_{i=0}^{n-1-\tau}s(i)s(i+\tau) $$ yܻANبCVk@IAM쥻ت|nA n |o n ӤnȡAoNO ACF uC $\tau=0$ ɡAACF |@ӳ̰IAoOڭ̭n䪺IC $\tau$ CCܤjɡAĤ@Ӱ򥻶g|MĤGӰ򥻶g|b@_A ACF S|X{ĤGӰIAoӰINOڭ̭n䪺IAIX{mANOڭ̭n䪺򥻶gC

ھڤWzAڭ̴NiH@qnTiءBp ACFBp⭵AöiӧX@qnVqAOѤFAROSA]٥pCӭتq]i²wqCӭؤTM^AYqӤpAhNت]wsANSC

ϥΤWz ACF IkANiH@qni歵lܡAdҦpUG

bWϤ@TӤpϡApUG

@쭵uAڭ̭nMƮwqiCMAƮwC@q]OƥনVqΦAq`ڭ̨ت׬O 32 msA]C| 1/32 = 31.25 ӭIAY@q 3 AVqN| 3*60*31.25 = 5625 ICӧڭ̪ۿJqnAYH 8 ҡAh| 8*31.25 = 250 ӭȡAڭ̪ؼСANOnXbo 5625 I̭A@q̹ڭ̰ۥXӪ 250 ӭȡC

Aڭ̻u̹vɡAoO@ӼҽkAҿסuvΡuvAھکڭ̩ҥΨ쪺ZơAZVphVAϤAhVCbpqVqZɡAڭ̥Ҽ{UCDG

Ĥ@ӰDAڭ̥iHNqP@ӭǡAAiCĤGӰDAڭ̥iH]tתܤƬOê]Ytק֡ANqY֨FYCANqYCAӤ|֩C^AbpUAڭ̴NiHĥΡuuʦYv]linear scalingA² LS^kӶiApUϩҥܡG

ѤWϤiHݥXAڭ̱NJVqԪ 1.25 ]PɱNJVqȥqVqȡ^ANMƮwY@qo̪񪺶ZAZYOJVqMqZC]bZƥiH²awqӦVqbתŶuZC^]b@qɡAڭ̥iHդPYơAҦpq 0.5B0.51B0.52B...B1.49B1.50 A@ 101 إiAӧX̨ΪYƥHι̵uZCYƮw 1000 qANo 1000 ӳ̵uZAڭ̥iھڳoǶZӱƧǡAZVuqANViOڭ̭۪qC

٦@ӰD٨SѨMGڭ̭nq̶}lOH@ӨAXإiG

ثeӷ~]ĥέۿqA̦WO SoundhoundAiHs http://www.soundhound.com iաCڭ̹Ǥ][F@ӭۿqiܨtΡA٬ MIRACLEAثeƮw 13000 qAĥ GPU iAɶu3Ai http://mirlab.org/demo/miracle ӶiաCHUOڸg MIRACLE ۤJu_Uv̫@yұo쪺GAѩu_UvӬO饻qu_ꤧKvAbUaQ½ۦUغqAMqWBqPAO۫٬O@ˡAҥHڭ̷|\h۫߻Pu_Uv@˪qA]tuGmKvBu_ꤧKvBuکMAvAoAHue˯vSʡGuݤeӤ meta dataC

ܦAڬ۫Hjaۿqwg@Ӱ򥻪AѡAγ\A̷|ݡAU@BOH٦޳ND|ݧJAHD٫ܦhAs]@biAHUCXXIG Ĥ@ӧxIAAi|ݡGOHճťoX MP3 ֤HnڡAqHAoO@ӤjvݡA٤OܩTaDHp󰵳oơAڭ̩TaDqoưnCC~@ӥ@ɪWQ| International Society of Music Information Retrieval]² ISMIR^||U˯A䤤@O audio melody extractionAMC~įೣW[Aثe raw pitch accuracy ٤ 85%AiHթMHTqF`ܦhCMաAq٬OHҳyXӪAҥHڭ̹qǮa̲ץؼСANOnyXMH@˼F`qAӭCHtAo˷||HhƩOHoNnaɶҩFI

i

  1. MIRǩҶ}o˯t
  2. ˯bΡGiܼv@iܼvG

Audio Signal Processing and Recognition (TBzP)