New labeling scheme with universal dictionary support and variance models with local retaking

[yqzhishen]

• New labeling system and phoneme duration predictor with universal dictionary support
• PitchDiffusion model for standalone, accurate and expressive pitch prediction
  • can be freely accelerated up to 100x or more without any loss of quality (several times faster than real-time on CPU)
  • predicts delta pitch from smoothed base pitch that is generated from note sequences
  • labels can be calibrated to get rid of out-of-tune errors even when training pitch data is not accurate
• MultiVarianceDiffusion model that jointly predicts all other variance parameters
  • predicts variance parameters like energy, breathiness, etc. based on phoneme information and pitch
  • Local retaking techniques that can do inpainting on a specific sub-region within pre-existing variance parameters