README.md

metadata

license: mit
language:
  - ja
pretty_name: ja005-22kh

removed data

seemed empty.VAD detect and remove

common_voice_ja_37417885
common_voice_ja_37498326
common_voice_ja_37501552
common_voice_ja_37517657

removed train_valid.txt

fasterwhisper detect and remove(still remain audios)

common_voice_ja_37409845
common_voice_ja_37417698
common_voice_ja_37454744
common_voice_ja_37500192
common_voice_ja_37500808
common_voice_ja_37500827
common_voice_ja_37506563
common_voice_ja_37515286
common_voice_ja_37521005
common_voice_ja_37515183
common_voice_ja_37535727
common_voice_ja_37628139

Mora Balance

OK?

{
'line-size':6750,'mora-size':172451,'min':2,'max':78,'mean':25,
moras:{
'a':3097,'i':9695,'u':2173,'e':3192,'o':8504,
'ba':1031,'bi':293,'bu':635,'be':379,'bo':314,
'bya':1,'byu':2,'bye':0,'byo':9,
'cha':93,'chi':1636,'chu':126,'che':9,'cho':276,
'da':1808,'di':9,'du':3,'de':3305,'do':1323,
'dya':0,'dyu':1,'dyo':0,
'fa':18,'fi':13,'fu':632,'fe':3,'fo':13,
'ga':3569,'gi':436,'gu':341,'ge':670,'go':566,
'gya':24,'gyu':8,'gye':0,'gyo':54,
'ha':963,'hi':854,'he':184,'ho':569,
'hya':13,'hyu':2,'hye':0,'hyo':97,
'ja':124,'ji':1581,'ju':396,'je':7,'jo':310,
'ka':6508,'ki':3958,'ku':4016,'ke':2209,'ko':3901,
'kya':119,'kyu':113,'kye':0,'kyo':391,
'ma':3062,'mi':1374,'mu':771,'me':946,'mo':3095,
'mya':1,'myu':3,'mye':0,'myo':20,
'na':5346,'ni':4196,'nu':110,'ne':545,'no':5511,
'nya':5,'nyu':37,'nye':0,'nyo':2,
'pa':260,'pi':49,'pu':127,'pe':53,'po':147,
'pya':1,'pyu':0,'pye':0,'pyo':7,
'ra':2929,'ri':2047,'ru':4492,'re':2776,'ro':871,
'rya':7,'ryu':19,'rye':0,'ryo':130,
'sa':1601,'si':0,'su':2425,'se':1691,'so':1525,
'sha':272,'shi':5596,'shu':438,'she':0,'sho':528,
'ta':5691,'ti':20,'tu':1,'te':5583,'to':5295,
'tsa':0,'tsi':0,'tsu':2589,'tse':1,'tso':0,
'tya':0,'tyu':0,'tyo':0,
'va':0,'vi':0,'vu':0,'ve':0,'vo':0,
'wa':4644,'wi':5,'we':3,'wo':1,
'ya':1003,'yu':351,'ye':0,'yo':1502,
'za':333,'zi':0,'zu':496,'ze':337,'zo':212
}
}

音声の品質

cer-kanji-0.085_kana-0.028_mer-0.012_per-0.012

by jyakoTen 0.2

日本語 Phonemize

とっても単純でシンプルなOpenJtalkによるg2p #注意事項あり

TrainとValidの分割 80:20

文字の長さと、音素のバランスV1を取った