yanggeng1995 / gan-tts Goto Github PK

Here:

Formula for hop_length : frame_shift_ms / 1000 * sample_rate instead of frame_shift_ms * 1000 / sample_rate . Same for win_length

Hi,

I was just wondering if you could maybe add a list of required python packages?
That would make setting up an environment for the TTS much easier.

Thanks in advance

请问有预处理过的数据集吗 源生的也行😂

hi，How to calculate the condition_window in dataset.py？ For example, if the sample_rate = 48000，hop_length = 600, what is the condition_window equals to? I will be looking forward to your reply，thank you.

Hi,
Thanks for sharing this nice implementation.

I can observe that you train the generator with a spectral (i.e. STFT-based) loss in addition to the adversarial one. While the original GAN-TTS paper trains the generator using only the adversarial one.

Have you tried any ablation study to check whether training with the adversarial loss only could give reliable reconstructions?

Also regarding the conditioning features: it seems that you use the Mel spectrogram for conditioning. Did you try using discrete linguistic features (i.e. embeddings of textual data)? Otherwise, what is the major difference between this model and the MelGAN one?

Many thanks in advance
Ahmed