Recap
2023-07-02 Hangzhou will be cloudy and sunny on Sunday
Note: I was born tone deaf, so I can never find the right tune when I sing. Let’s use AI to synthesize timbre. It has only been trained for about 15,000 steps so far. The data set I prepared is actually only about 40 minutes long. For data preprocessing, It has been deleted a bit, so there is not a lot of sound data, but after all I have been busy all afternoon, I have to take a look at the effect. I plan to train about 30,000 to 50,000 steps before stopping. If it is more, my disk may be damaged. That’s not possible anymore.
AI Reasoning (Synthesis)
1. Inference parameters
2. Audio conversion
3. Audio synthesis
4. Test music
NetEase Cloud Music (continuously updated): https://music.163.com/#/outchain/4/991811271/
Summarize
I finally synthesized the first song. After training for almost 4 hours, I still have to continue. 12,000 steps still cannot achieve the expected results. I turned on the computer to make elixirs this week. When training was relatively small, my voice was obviously hoarse. , but the training audio is clear in articulation. I hope that after two days of training, it can be directly used to synthesize AI songs.