VoiceMaker-1.1 ― HMM音声合成用音響モデルの構築

Size: px

Start display at page:

Download "VoiceMaker-1.1 ― HMM音声合成用音響モデルの構築"

やすはるねごろ
5 years ago
Views:

1 VoiceMaker-1.1 HMM / sako@mmsp.nitech.ac.jp

2 2 VoiceMaker-1.1 HMM 1 Galatea Toolkit (GalateaTalk) HTS *1 GalateaTalk HTS 1.1 Linux PC CPU CPU MB 503?? HTK HMM HTS HMM HTK SPTK Julius get f0s F 0 1 HMM *1

3 3 HTS ATR 503 HMM ATR *2 ATR MHT F HMM HTS *3 HTS HMM HTK (Hidden Markov models Tool Kit) *4 HTK HTK HMM HTS HTK SPTK (Signal Processing Tool Kit) * Julius *6 Julius segment.pl HMM *2 HTS www *3 *4 *5 tokuda/sptk/ *6

4 4 VoiceMaker-1.1 HMM Galatea Talk Galatea Toolkit SPTK HTK HTS 2 README config_ini.pl mkdata.pl configure Makefile.in tools data detect_f0range.sh record.pl segment.pl Makefile.in scripts questions 3 F 0 F 0 get f0s Snack *7 Wavesurfer *8 ESPS get f0 F 0 2 VoiceMaker VoiceMaker-1.1.tar.gz 500 1GB 3 % tar zxf VoiceMaker-1.1.tar.gz % cd VoiceMaker config.pl Perl *7 *8

5 5 base config.pl speaker dataset atr 503 ATR 503 istc b50 ISTC 50 recdir datadir sptkbindir SPTK README htsbindir HTS README julius dickit Julius dictation kit getf0 base F 0 mceporder * 9 3 atr ( ) PC 16kHz USB (record.pl) Perl r please speak. 16kHz 16bit short Big./speech/ %./record.pl [a01] * fragment size = 1024 bytes (32msec) AD-in thread created <<< please speak >>... % *9 gtalk

6 6 VoiceMaker-1.1 HMM l r t b n m X X ( m b01) q play sox % play -x -r f s -s w -c 1.ad % sox -r c 1 -s -w -x input.ad output.wav Microsoft WAV % play.wav HMM HTK SPTK HTK MFCC LPC (HCopy) 4.1 F 0 Julius./data/

7 7 Hz a r a y u r u g e N j i ts u o [msec] 4 :... %./mkdata.pl 4.2 Julius 4 HTK (HSLab) wavesurfer./data/labels/mono/ HTK 100ns sil a r a y u r u g e : : wavesurfer wavesurfer Choose Configuration HTK transcription Load Transcription..

8 8 VoiceMaker-1.1 HMM./data/labels/mono/ 4.3 F 0 ( Hz 0) -U/-L % /opt/get_f0s/get_f0s -z -Z 3 -U 400 -L 60 -r b raw/istc_balance_b01.raw \ > f0/istc_balance_b01.f0 * 10 F 0./data/f0/ ( Hz) (float) SPTK F 0 % PATH=$PATH: /opt/sptk-3.0/bin (SPTK ) % gwave -i 1 +s f0/.f0 xgr 4.4 HMM SPTK make * 11 % make data HTK./data/cmp/ HMM SPTK./data/mcep/ *10 Wavesurfer *11 mkdata.pl Makefile

9 9 x2x short float frame window mcep 4.6 HTS 18 2 HTK HMM * 12 HTK HList * 13 % HList -C work/?/data/hlist.conf work/?/data/cmp/istc_?_b01.cmp 0: e10-1.0e10-1.0e10 1: e10-1.0e10-1.0e10 2: e10-1.0e10-1.0e10 : : 4.7 make HTS lists/mono?.listg lists/full?.list (./cmp/*.cmp) lists/data istc b50?.scp label/mono?.mlf, label/full?.mlf *12 HTK 12 *13-1.0e10

10 10 VoiceMaker-1.1 HMM 2 speech signal 2 HTK 5 HInit, HRest HMM HHEd HMM HERest HERest HHEd HHEd HERest HHEd 6 5 HMM 6

11 make scripts/training *.pl gtalk 450 (Pentium4 3GHz PC ) MB % cd.. % perl./scripts/training.pl scripts/config.pl Makefile SPTKBINDIR SPTK HTSBIN HTS DATADIR NAME ( ) DATASET ( ) MCEPORDER HMM HTK HInit HMM HRest HERest 5.3 HMM HMM HMM HMM { } * 14 { } { } { } *14

12 12 VoiceMaker-1.1 HMM 1 p L p C p R a C1 a C2 b L1 b L2 b L3 b C1 b C2 b C3 b R1 b R2 b R3 c L1 c L2 c L3 c L4 c C1 c C2 c C3 c C4 c C5 c R1 c R2 c R3 c R4 d L1 d C1 d C2 d R1 e ( : ) ( : ) ID ID ID ID ID ID ID ID ID ( : ) ( : ) ( : ) ( : ) ( : ) ( : ) ( : ) a 1./labels/full/ p L -p C +p R /A:a C1 a C2 /B:b L1 b L2 b L3 -b C1 b C2 b C3 +b R1 b R2 b R3 /C:c L1 c L2 c L3 c L4 -c C1 c C2 c C3 c C4 c C5 +c R1 c R2 c R3 c R4 /D:d L1 -d C1 d C2 +d R1 /E:e * 15 HHEd HMM

13 13 HERest 5.4 HTK HHEd./questions/question.hed : QS R_yusei_masatuon { *+j/a*,*+z/a* } QS R_musei_masatuon { *+fy/a*,*+hy/a*,*+sh/a*,*+s/a*,*+f/a*,*+h/a* } QS R_yusei_haretuon { *+dy/a*,*+gy/a*,*+vy/a*,*+by/a*,*+g/a*,*+d/a*,*+v/a*,*+b/a* } QS R_musei_haretuon { *+ky/a*,*+py/a*,*+k/a*,*+t/a*,*+p/a* } : 7 * 16 6 HTS TTS (Text-to-speech) GalateaTalk *

14 14 VoiceMaker-1.1 HMM C_yuseion L_yuseion no yes no yes sub_accentgata_mora<=-6 mora==x no yes no yes mcep_s2_1 mcep_s2_2 C_masatsuon L_ago_kou_boin C_k no yes mcep_s2_5 no mcep_s2_6 yes mcep_s2_7 no yes mcep_s2_3 mcep_s2_ S-1 4byte d 1 d N N 8 GalateaTalk 6.1 GalateaTalk GalateaTalk HTK GalateaTalk HTS HHEd./voices/.pdf,

15 15.inf float 8 HHEd./edfiles/?/?/cmp/convert? qst001.hed 6.2 GalateaTalk./voices/?/?/qst001/ 6 GalateaTalk gtalk ssm.conf -SPEAKER-ID TREE-FILE -MODEL-FILE SPEAKER-ID: test GENDER: male DUR-TREE-FILE: / /tree-dur.inf PIT-TREE-FILE: / /tree-lf0.inf MCEP-TREE-FILE: / /tree-mcep.inf DUR-MODEL-FILE: / /duration.pdf PIT-MODEL-FILE: / /lf0.pdf MCEP-MODEL-FILE: / /mcep.pdf gtalk gtalk set Speaker = test 7 GalateaTalk HMM GalateaTalk 8 : config.pl /opt/ /opt Vine Linux SPTK SPTK SPTK SPTK tokuda/sptk/ SPTK-3.0.tar.gz (PDF )

16 16 VoiceMaker-1.1 HMM /opt/ % cd /opt % tar zxf SPTK-3.0.tar.gz () % cd SPTK-3.0/src % make PREFIX=/opt/SPTK-3.0 ( /usr/local/sptk ) % make install /opt/sptk-3.0/bin/ SPTK 8.2 get f0s ESPS get f0 get f0s % cd /opt % tar zxf get_f0s-0.1.tar.gz % cd get_f0s-0.1 % make 8.3 Julius Julius Linux Julius dictation-kit-v3.1.tar.gz % cd /opt % tar zxf dictation-kit-v3.1.tar.gz () 8.4 HTS HTS HTK(HMM Tool Kit) HTK HTK HTS HTK HTK-3.4

17 17 HTS-2.0 /opt/htk/ HTK HTK-3.4.tar.gz, HDecode.tar.gz HTS for HTK-3.4.tar.gz HTK HTS % cd /opt % tar zxf HTK-3.4.tar.gz () % tar zxf HDecode-3.4.tar.gz () % cd htk % tar zxf HTS-2.0_for_HTK-3.4.tar.gz () HTS HTS-2.0 for HTK-3.4.patch HTK-3.4 % patch -p0 < HTS/HTS-2.0_for_HTK-3.4.patch HTK HTS HTK make %./configure --prefix=/opt/htk % make % make install /opt/bin/ HTS * 17 *17 HTK

2

2 6 7 9 4 6 7 2 3 4 5 6 7 8-0 - G G G G G G f f 9 e f - e f 0 5 e fe e c c cc B FD F 5 2 5 D F C e e e b 3 f f 5 ff ff f f f f b b bb b b b c c c ee ee e ee ee e f f 4 e e 7 5 5 e bb 6 7 f GE 8 f 9 5 F