Description

After a few days of rest, I also did some things~ put it together~

All tasks and DDL

[Old School] SLT paper reading and discussion
[Company] [Ivory Tower] PPG Ali paper version. The tacotron part, draw the localization algorithm. Draw the form to design the hyperparameter search experiment. Discuss using rayhane or torch. PPG extraction and write the algorithm, by the way, train the cpu version, and learn Source PPT. After the SLT, set the time again , and expect a preliminary result before January 31
[School] [Company] Mixed language stage2, bilingual corpus LID modified synthesis results sorting. Set time after SLT
[Company] [Ivory Tower] Summarized the experimental results of the remaining three types of hyperparameters . Under the design λ, the effect of dim and wide on the VC capability. similar loss framework). Set the time after SLT
[School] Summarize the PPG Project method as a patent. The content of the PPT is organized into a blog to sort out some things. After the SLT, set the time
[Company] Maybe-TTS is transferred according to the module, optimized and merged to calculate the py file of the pause text
[Old School] Sort out Xintao's papers, the teacher recommended to find the latest paper for reading and sorting, connect with Yelaoshr’s melgan, and consider the Tacotron down-mixed text processing plan based on the notes. Read the paper class. Look at the ML book. Take your time

[School] [ivory tower] Show the accented paper PPT and read the paper again, plus the parallel corpus of the data set. I went to the laboratory to find the source at 8:00 am on 1/19. The answer is very good~, it's over
[Company] [Ivory Tower] AutoVC (sil, dataloader, val cutting), look at the results, re-adjust parameters, one-hot version, hyperparameter comparison. Add the output during training to wav, val and conversion. Do not use wavenet as output , Write GL. Use the neural network to do it again, add a new version, improve the blog, analyze, and give the experience and skills of AutoVC
~~Discussed syntax tree related content with Changhe, and Tree LSTM~~