PITS

PITS: Variational Pitch Inference without Fundamental Frequency for End-to-End Pitch-controllable TTS

Abstract: Previous pitch-controllable text-to-speech (TTS) models rely on directly modeling fundamental frequency, leading to low variance in synthesized speech. To address this issue, we propose PITS, an end-to-end pitch-controllable TTS model that utilizes variational inference to model pitch. Based on VITS, PITS incorporates the Yingram encoder, the Yingram decoder, and adversarial training of pitch-shifted synthesis to achieve pitch-controllability. Experiments demonstrate that PITS generates high-quality speech that is indistinguishable from ground truth speech and has high pitch-controllability without quality degradation. Code and audio samples will be available at https://github.com/anonymous-pits/pits.

Training code is uploaded.

Demo and Checkpoint are uploaded at Hugging Face Space🤗

Audio samples are uploaded at github.io.

For the pitch-shifted Inference, we unify to use the notation in scope-shift, s, instead of pitch-shift.

Voice conversion samples are uploaded.

Accepted to ICML 2023 Workshop on Structured Probabilistic Inference & Generative Modeling

Requiremetns

2.0>Pytorch>=1.7.0 for alias-free-torch and phaseaug. Since our codes are based on lower version of torch, do not use 2.x.
The requirements are highlighted in requirements.txt.
We also provide docker setup Dockerfile.

docker build -t=pits .

For VCTK

Download VCTK (version 0.92) at link
Resample files to 22050 Hz, 16 bit, .wav format.

For custom dataset

Do not train this model with single speaker dataset, due to our pitch-shift training applying GAN, it seems that single speaker dataset cannot trained with it. Many issues are reporting training failure for small or single speaker dataset.
We strongly encourage to use pretrained models with speaker embedding size editing. It requires long training time (over 3 weaks with 4 V100 GPUs), so you should apply it for efficient traning.
Resample files to 22050 Hz, 16 bit, .wav format. Some issues are reporting training failures for other sampling rate, we do not ensure for other sampling rates.
Check configs/config_en.yaml and change data section (e.g. path, files, speakers).
To utilize a different phoneme set, you will need to modify the Python files located in the text folder.
To train other languages than English, please refer other language variants of VITS. There are many references in github.

Training

Run train.py, check train.py for detailed options.

python train.py -c configs/config_en.yaml -m {MODEL_NAME} {-i:if you change yingram setup or etc}

Demo