|
Yes, I installed the SPTK-3.5, My training data have 4000 utterance, I do not checked any one of the training data, but I sample 100 utt, and all of them have the same problem that the mgc file have less two frames than the raw file. the raw files are 16k sample rate, 16bit pcm, little endian data format. the details command below: FRAMELEN = 400 FRAMESHIFT=80 FFTLEN=512 WINDOWTYPE =1 FREQWARP=0.42 GAMMA=0 MGCORDER=20 > Date: Thu, 14 Jun 2012 00:55:28 +0900 > From: uratec@xxxxxxxxxxxxxxx > Subject: [hts-users:03347] Re: About getf0 and cmp data file？ > To: hts-users@xxxxxxxxxxxxxxx > CC: uratec@xxxxxxxxxxxx > > Hi, > > Do you use SPTK-3.5? > Anyway, please let me know the number of *samples* of your raw file. > > Regards, > Keiichiro Oura > > > 2 012/6/9 ArthurLeo <bin007.zhao@xxxxxxxxxxx>: > > Thank you very much for detail answer, here is the further question： > > 1. Now assuming the F0 is 200Hz, acctually, the f0 can be varied from 60Hz > > to 400Hz generally, so even if 5ms waveform added to the head, the first > > frame also can not be located the start time of the raw file certainly. > > is it? > > > > 2. From the attached illustration, the mgc file should have the same frames > > as the raw file, but in my test, the mgc file have less two frames than the > > raw file, > > is it abnormal? > > > >> Date: Sat, 9 Jun 2012 17:43:40 +0900 > >> From: tokuda@xxxxxxxxxxxx > >> Subject: [hts-users:03344] Re: [hts-users:03342] RE: [hts-users:03341] Re: > >> About getf0 and cmp data file？ > >> To: hts-users@xxxxxxxxxxxxxxx > >> CC: tokuda@nit ech.ac.jp > >> > >> > 1. yes, I guess, but I am not sure about the snack a lgorithm of getting > > > >> > pitch, so I wander to know why the 5ms/25ms, not other length? > >> > >> We add 5ms to the file head in order to adjust the positions of "frame > >> center" > >> assuming the extracted F0 is always 200Hz. > >> > >> See the attached illustration. > >> > >> Keiichi > >> > >> > >> 2012/6/9 ArthurLeo <bin007.zhao@xxxxxxxxxxx>: > >> > 1. yes, I guess, but I am not sure about the snack algorithm of getting > >> > pitch, so I wander to know why the 5ms/25ms, not other length? > >> > > >> > 2. My raw file have 3104 frames，16k sample rate, FRAMELEN = > >> > 400 FRAMESHIFT=80 FFTLEN=512 WINDOWTYPE =1 FREQWARP=0.42 GAMMA= 0 > >> > MGCORDER=20, but after mgc analysis, the mgc file have less two frames > >> > (3102 > >> > frames) , and the lf0 file have the same number frames（3104） with the > >> > original speech raw file , maybe because of the added head/tail. when > >> > the< br>> > stage using the sptk tool merge function to compose the cmp > >> > file，mgc file > > > >> > inserted > >> > the lf0 file，so the final cmp file have the same frames with mgc file > >> > not > >> > the lf0 file，and lf0 file two more frames at the tail were discarded > >> > directly. > >> > So I was confused, and want to know the detail reason. is it reasonable? > >> >> Date: Sat, 9 Jun 2012 00:08:44 +0900 > >> >> From: uratec@xxxxxxxxxxxxxxx > >> >> Subject: [h ts-users:03341] Re: About getf0 and cmp data file？ > >> > > >> >> To: hts-users@xxxxxxxxxxxxxxx > >> >> CC: uratec@xxxxxxxxxxxx > >> >> > >> >> Hi, > >> >> > >> >> > 1. Why added 5ms/25ms waveform to head/tail? why not other length, > >> >> > like > >> >> > 10ms head or the 5ms tail? I guess it must have some special reason, > >> >> > I > >> >> > read > >> >> > the getf0.tcl. the default minpitch/maxpitch for speech is 60/400Hz, > >> >> > so if the autocorrelation method is generally used to caculate the > >> >> > first > >> >> > frame pitch, it should be added (1/60)s=16.7ms waveform for the head > >> >> > at > >> >> > least, but now the numbe r is 5ms/25ms, so can you explain more > >> >> > detail? > >> >> > >> >> Do you mean that the lower F0 limit affect the number of generated F0 > >> >> frames? > >> >> When I changed the lower F0 limit, the number of generated F0 frames > >> >> were not changed. > >> >> > >> >> > 2. I checked the *.mgc file, it always have less two frames than the > >> >> > *.raw > >> >> > file, so my question is the discarded two frames is the raw file head > >> >> > two > >> >> > frame or the raw file tail frame ? > >> >> > >> >> It's strange. > >> >> Please, let me know the number of samp les of your raw file, command > > > >> >> lines of mgc analysis, and the number of generated mgc frames, > >> >> respectively. > >> >> > >> >> Regards, > >> >> Keiichiro Oura > >> >> > >> >> > >> >> 2012/6/8 ArthurLeo <bin007.zhao@xxxxxxxxxxx>: > >> >> > Thank you for your answer, but I have some further questions about > >> >> > the > >> >> > details of the lf0 and mgc parameter extraction. > >> >> > 1. Why added 5ms/25ms waveform to head/tail? why not other length, > >> >> > like > >> >> > 10ms head or the 5ms tail? I guess it must have some special reason, > >> >> > I > >> >> > read > >> >> > the getf0.tcl. the default minpitch/maxpitch for speech is 60/400Hz, > >> >> > so if the autocorrelation method is generally used to cacu late the > >> >> > first > >> >> > frame pitch, it should be added (1/60)s=16.7ms waveform for the head > >> >> > at > >> > > > least, but now the number is 5ms/25ms, so can you explain more > >> > > > detail? > >> >> > > >> >> > 2. I checked the *.mgc file, it always have less two frames than the > >> >> > *.raw > >> >> > file, so my question is the discarded two frames is the raw file head > >> >> > two > >> >> > frame or the raw file tail frame ? > >> >> > > >> >> > > ; Date: Fri, 8 Jun 2012 09:01:56 +0900 > >> >> >> From: uratec@xxxxxxxxxxxxxxx > >> >> >> Subject: [hts-users:03339] Re: About getf0 and cmp data file？ > >> >> >> ; To: hts-users@xxxxxxxxxxxxxxx > >> >> >> CC: uratec@xxxxxxxxxxxx > >> >> > > >> >> >> > >> >> >> Hi, > >> >> >> > >> >> >> The number of frames generated by Snack (in ActiveTcl) is often > >> >> >> lower > >> >> >> than the number of frames generated by SPTK. > >> >> &g t;> The internal frame calculation is different between Snack and > >> >> SPTK. > >> > > >> >> >> Therefore, 5ms/25ms waveform are added to head/tail of the utterance > >> >> >> before f0 analysis in the HTS demo script. > >> >> >> > >> >> >> Regards, > >> >> >> Keiichiro Oura > >> >> >> > >> >> >>> >> >> >> 2012/6/7 ArthurLeo <bin007.zhao@xxxxxxxxxxx>: > >> >> >> > Hi all， > >> >> >> > When I prepared the cmp data for training， I found that the *.lf0 > >> >> >> > file > >> >> >> > always have two more frames than the *.mgc file, and they are > >> >> >> > extracted > >> >> >> > from > >> >> >> > the same *.raw file. So I checked the ./data/makefile，and I foun d > >> >> & gt;> > the > > > >> >> >> > *.raw > >> >> > > >> >> >> > file were added the 0.5ms head data and the 25ms tail data，I guess > >> >> >> > it > >> >> >> > maybe > >> >> >> > the reason，but I don not know why do this？why add the head and t > >> >> >> > he > >> >> >> > tail？ > >> > > >> >> >> > why > >> >> >> > 0.5ms and 25ms，not others？ Can some one known about it？ when > >> >> >> > composing > >> >> >> > the > >> >> >> > lf0 and mgc with the SPTK tool merge function，how deal with the > >> >> >> > the > >> >> >> > extra > >> >> >> > lf0 frame？ discarded directly？ > >> >> >> > >> >> > |