Implementation of TTS with combination of Tacotron2 and HiFi-GAN for Mandarin TTS.


In order to inference, we need to download pre-trained tacotraon2 model for mandarin, and place in the root path. Then, we can run infer_tacotron2_hifigan.py to get TTS result. We can alter the input text by editting variablle text in the infer_tacotron2_hifigan.py. Then the result will be saved in the root path named as output.wav.

The pre-trained model of HiFi-GAN has been placed in the LJ_FT_T2_V3, which is trained by LJSppech and fine-tuned with Tacotron2. You can find more pre-trained model from original HiFi-GAN repo with different size and parameters. If you want to try different models or train your own model, please do remember to alter variables in infer_tacotron2_hifigan.py to change the path of HiFi-GAN model.

Audio Sample

Input: 相对论直接和间接的催生了量子力学的诞生 也为研究微观世界的高速运动确立了全新的数学模型
Output: tacotron2-hifigan.wav


View Github