carpedm20 / multi-speaker-tacotron-tensorflow

Multi-speaker Tacotron in TensorFlow.
http://carpedm20.github.io/tacotron
Other
632 stars 341 forks source link

사용자 지정 Datasets으로 학습을 진행하는데 문제가 있는것같습니다 #36

Closed rycont closed 5 years ago

rycont commented 6 years ago

test-step-000016000-align001 16000Step학습한 testStep 그래프입니다. 현재 '하ㅈ......'정도가 들리는 상태이며 synthesizer.py로 만들어낸 결과물은 alignment.json에 있는 내용을 제외하면 거의 괴음이 들립니다. alignment.json에 있는 문장도 만들어보면 첫 자음정도까지만 들리고 괴음이 들립니다. 학습 데이터 양의 문제가 있는걸까요? test-step-000016000-audio001.zip 16000step testStep 결과물 첨부해봅니다

rycont commented 5 years ago

자답입니다. 학습 데이터 양 부족입니다.