Add proof of concept shallow AR model implementation

[r9y9]

Fixes #15

Summary

This PR adds support for shallow AR models which was proposed in https://www.researchgate.net/publication/324609423_Autoregressive_neural_F0_model_for_statistical_parametric_speech_synthesis.

There are several differences from the original paper (e.g., continuous vs. quantized f0, joint feature modeling vs. F0 only). However, I think I have implemented the core idea.

How it works

The following figure shows a comparison of spectrograms between 1) the previous model (top), 2) the shallow AR model (bottom). As you can see, the shallow AR model is better at capturing the time-varying nature of F0. Vibrato is well modeled without explicit modeling of vibrato parameters.

Samples

1. The previous model: https://soundcloud.com/r9y9/20200522-haru-ga-kita-3-nit-song070?in=r9y9/sets/dnn-based-singing-voice
2. Shallow AR model: https://soundcloud.com/r9y9/20200522-haru-ga-kita-4-nit-song070-sar-test?in=r9y9/sets/dnn-based-singing-voice

Resouces

• https://tonywangx.github.io/research.html
• https://www.research.ed.ac.uk/portal/files/65079598/Autoregressive_neural_F0_model_for_statistical_parametric_speech_synthesis.pdf

[r9y9]

I've added a SAR recipe that does not use MLPG. New samples were uploaded:

1. svs-world-conv: https://soundcloud.com/r9y9/2020117-yuki-nit-song070-svs-world-conv
2. svs-world-conv-sar: https://soundcloud.com/r9y9/2020117-yuki-nit-song070-svs-world-conv-sar
3. svs-world-conv-sar-wo-mlpg: https://soundcloud.com/r9y9/2020117-yuki-nit-song070-svs-world-conv-sar-wo-mlpg

svs-world-conv-sar-wo-mlpg works okay, but the pitch seems like pitch is less stable than the models with MLPG.

[r9y9]

Based on my experiments, shallow AR models works better than the previous model for the NIT SONG 070 dataset. However, I couldn't find good configurations for the other dataset yet. I will leave it for future work.

Let me merge the PR, and let's continue testing on different dataset and configurations.