search

sh-lee-prml / HierSpeechpp

The official implementation of HierSpeech++
MIT License
1.17k stars 134 forks source link

Seemingly incorrect tts result #16

Open MaestroT opened 9 months ago

MaestroT commented 9 months ago

Hi, I'm unfamiliar with Korean, while I got difference output from your work and a tts website with input text "이전 곡". I generated audio from 4 different ways: 1: using inference with code from this repo and checkpoint logs/hierspeechpp_eng_kor/hierspeechpp_v1.1_ckpt.pth 2: using your huggingface demo page: https://huggingface.co/spaces/LeeSangHoon/HierSpeech_TTS 3: using coqui ai's huggingface page: https://huggingface.co/spaces/coqui/xtts 4: using 2 free tts websites: https://www.text-to-speech.cn/ and https://ttstool.com/ Here's what I got: https://drive.google.com/drive/folders/1BfptrylJTmICm2JN49G2YQi-HjnOSCBB?usp=sharing

sh-lee-prml commented 9 months ago

Hi!

We do not utilize Korean dataset to train the TTV model so current TTS system only supports English! (Only speech synthesizer was trained with English and Korean dataset!)

As we use a Phonemizer, Korean text could be fed to the model but the speech will not be synthesized properly.

Thanks!

MaestroT commented 9 months ago

Hi!

We do not utilize Korean dataset to train the TTV model so current TTS system only supports English! (Only speech synthesizer was trained with English and Korean dataset!)

As we use a Phonemizer, Korean text could be fed to the model but the speech will not be synthesized properly.

Thanks!

Thanks! Will you then conduct training for Korean TTS?

sh-lee-prml commented 9 months ago

Sorry for the late reply.

We do not have enough GPUs for training the Korean TTS so we postponed training the Korean TTS now...

Actually, I'm currently training new model for other paper... I appreciate your understanding.

Thanks!