首页 > 其他分享> > 基于 RTVC-7 Voice Cloning Model 的 Cross-Lingual TTS 的三步走: 第二步 Tuned-GE2E-EarSpeech 不跨语言实验观察
基于 RTVC-7 Voice Cloning Model 的 Cross-Lingual TTS 的三步走: 第二步 Tuned-GE2E-EarSpeech 不跨语言实验观察
作者:互联网
0. 说明
- 训练的非常好
- 进行实验的设置和张阳 Speaker Embedding 的一样: https://blog.csdn.net/u013625492/article/details/114640032
- 代码逻辑从: /ceph/home/hujk17/Tuned-GE2E-EarSpeech/FaPig_GE2E_synthesizer.py 开始
1. 选取 Reference Speech
使用里面的 speaker embedding 的路径, 然后替换为 GE2E 的路径, 其实就多了个后缀, 参与网络预测, 然后用 gl 合成里面的 mel, 作为参考
- /ceph/home/hujk17/npy-EarSpeech-HCSI-Data/dereverb_npy/SSB0005/mel-SSB00050119-mel.npy|/ceph/home/hujk17/npy-EarSpeech-HCSI-Data/dereverb_npy/SSB0005/spk-SSB00050119.npy|81000|406|ti3 yu4 zong3 ju2 gai1 gan4 de5 rang4 ti3 yu4 zong3 ju2 gan4
- /ceph/home/hujk17/npy-EarSpeech-HCSI-Data/dereverb_npy/SSB0005/mel-SSB00050117-mel.npy|/ceph/home/hujk17/npy-EarSpeech-HCSI-Data/dereverb_npy/SSB0005/spk-SSB00050117.npy|97800|490|you3 liu4 shi2 jia1 gong1 si1 wei4 xin1 xing1 chan3 ye4 shang4 shi4 gong1 si1
- /ceph/home/hujk17/npy-EarSpeech-HCSI-Data/dereverb_npy/SSB0323/mel-SSB03230294-mel.npy|/ceph/home/hujk17/npy-EarSpeech-HCSI-Data/dereverb_npy/SSB0323/spk-SSB03230294.npy|80000|401|xing2 cheng2 ji4 you3 fen1 gong1 you4 xiang1 hu4 pei4 he2 de5 jian1 guan3 ji1 zhi4
- /ceph/home/hujk17/npy-EarSpeech-HCSI-Data/dereverb_npy/SSB0323/mel-SSB03230295-mel.npy|/ceph/home/hujk17/npy-EarSpeech-HCSI-Data/dereverb_npy/SSB0323/spk-SSB03230295.npy|57200|287|ping2 jun1 shou4 jia4 si4 bai2 jiu3 shi2 jiu2 mei3 yuan2
- /ceph/home/hujk17/npy-EarSpeech-HCSI-Data/dereverb_npy/SSB1711/mel-SSB17110171-mel.npy|/ceph/home/hujk17/npy-EarSpeech-HCSI-Data/dereverb_npy/SSB1711/spk-SSB17110171.npy|83400|418|nan2 guai4 ta1 zi4 chao2 dan1 xin1 zhe4 ge4 yang4 zi5 xing2 chu1 jie1 hui4 bei4 ren2 da3
- /ceph/home/hujk17/npy-EarSpeech-HCSI-Data/dereverb_npy/SSB1711/mel-SSB17110174-mel.npy|/ceph/home/hujk17/npy-EarSpeech-HCSI-Data/dereverb_npy/SSB1711/spk-SSB17110174.npy|80600|404|qi1 huo4 ying2 ye4 bu4 shu4 liang4 shi2 xian4 le5 ping2 wen2 you3 xu4 zeng1 zhang3
- /ceph/home/hujk17/npy-EarSpeech-HCSI-Data/tst_npy/MST-Originbeat-S2-male-5000/mel-002126-mel.npy|/ceph/home/hujk17/npy-EarSpeech-HCSI-Data/tst_npy/MST-Originbeat-S2-male-5000/spk-002126.npy|79200|397|lao3 bo2 jue2 shuo1 ge1 zi5 nen2 zai4 ta3 shang5 xie1 xi5 na4 qun2 huo3 ji1 ze2 zhu4 zai4 di4 yi1 cen2
- /ceph/home/hujk17/npy-EarSpeech-HCSI-Data/tst_npy/MST-Originbeat-S2-male-5000/mel-002131-mel.npy|/ceph/home/hujk17/npy-EarSpeech-HCSI-Data/tst_npy/MST-Originbeat-S2-male-5000/spk-002131.npy|48400|243|san1 liang4 yue4 ye3 che1 zai4 cong2 lin2 dao4 lu4 shang5 gao1 su4 lue4 guo4
- /ceph/home/hujk17/npy-EarSpeech-HCSI-Data/dereverb_npy/SSB1711/mel-SSB17110187-mel.npy|/ceph/home/hujk17/npy-EarSpeech-HCSI-Data/dereverb_npy/SSB1711/spk-SSB17110187.npy|59600|299|er2 wo3 men5 zhe4 ge4 tuan2 dui4 jin1 nian2 xin1 zu3 cheng2
- /ceph/home/hujk17/npy-EarSpeech-HCSI-Data/dereverb_npy/SSB1711/mel-SSB17110188-mel.npy|/ceph/home/hujk17/npy-EarSpeech-HCSI-Data/dereverb_npy/SSB1711/spk-SSB17110188.npy|61400|308|gao1 dang4 jiao4 che1 xu1 jia1 jiu3 shi2 qi1 hao4 qi4 you2
- /ceph/home/hujk17/npy-EarSpeech-HCSI-Data/tst_npy/MST-Originbeat-S1-female-5000/mel-000196-mel.npy|/ceph/home/hujk17/npy-EarSpeech-HCSI-Data/tst_npy/MST-Originbeat-S1-female-5000/spk-000196.npy|42600|214|si1 ding1 ni2 ba3 suo2 you3 zhe4 yi2 qie4 ao2 zai4 guo1 li3
- /ceph/home/hujk17/npy-EarSpeech-HCSI-Data/tst_npy/MST-Originbeat-S1-female-5000/mel-000456-mel.npy|/ceph/home/hujk17/npy-EarSpeech-HCSI-Data/tst_npy/MST-Originbeat-S1-female-5000/spk-000456.npy|79400|398|jia2 dao3 zai4 ye3 bu2 yong4 wei4 ying2 zhan3 qiu2 ren2 le5 ye3 zhong1 yu2 ke2 yi3 shui4 jiao4 le5 wo3 hai2 ting3 kai1 xin1 de5
- /ceph/home/hujk17/npy-EarSpeech-HCSI-Data/tst_npy/MST-Originbeat-S1-female-5000/mel-004726-mel.npy|/ceph/home/hujk17/npy-EarSpeech-HCSI-Data/tst_npy/MST-Originbeat-S1-female-5000/spk-004726.npy|52600|264|zhun3 bei4 zhi1 fu4 yi4 qian1 wan4 zui4 duo1 yi4 qian1 wu2 bai3 wan4 mei3 yuan2
- /ceph/home/hujk17/npy-EarSpeech-HCSI-Data/tst_npy/MST-Originbeat-S2-male-5000/mel-001556-mel.npy|/ceph/home/hujk17/npy-EarSpeech-HCSI-Data/tst_npy/MST-Originbeat-S2-male-5000/spk-001556.npy|89000|446|shi4 bing1 shuo1 dao4 ta1 de5 zui2 jiao3 lu4 chu1 le5 yi4 si1 jian1 xiao4 ta1 xing1 li3 fei1 chang2 qing1 chu3 zhe4 jian4 shi4
- /ceph/home/hujk17/npy-EarSpeech-HCSI-Data/tst_npy/MST-Originbeat-S2-male-5000/mel-001576-mel.npy|/ceph/home/hujk17/npy-EarSpeech-HCSI-Data/tst_npy/MST-Originbeat-S2-male-5000/spk-001576.npy|56000|281|hai2 you3 wang2 you3 wen4 ai4 si1 bo2 ni3 wei4 shen2 me5 zhe4 me5 pa4 zhong1 guo2
2. 选取文本
- ti3 yu4 zong3 ju2 gai1 gan4 de5 rang4 ti3 yu4 zong3 ju2 gan4
- you3 liu4 shi2 jia1 gong1 si1 wei4 xin1 xing1 chan3 ye4 shang4 shi4 gong1 si1
- xing2 cheng2 ji4 you3 fen1 gong1 you4 xiang1 hu4 pei4 he2 de5 jian1 guan3 ji1 zhi4
- wo3 xi3 huan1 ni3 zhou1 cui4 bao4 zao4 ge1
- xie4 xie4 da4 jia1 lin2 ting1 yu3 yin1 he2 cheng2 shi2 yan4
3. Speaker 相似的具体现象
3.1. 急促感
当参考音频比较急促时: speaker id = 2, GE2E 在句末更短促, 效果更好
标签:GE2E,TTS,mel,hujk17,Tuned,ceph,EarSpeech,npy,home 来源: https://blog.csdn.net/u013625492/article/details/114658654