Lower quality voice synthesize with the new version

I trained a speech model and the vocoder using the new version (espnet==0.10.6 and parallel_wavegan==0.5.4) and the synthesized voices have low quality. But when I use the old espnet (espnet==0.9.6 and parallel_wavegan==0.4.8) for inference the outputs have good quality. The new code for inference looks like this (the version that doesn’t work for me):

tts = Text2Speech.from_pretrained(model_file="exp/tts_train_conformer_fastspeech2_raw_phn_tacotron_g2p_en_no_space/1000epoch.pth",
                                  vocoder_file="/home/rohola/codes/ryan-tts/models/vocoders/train_nodev_ryanspeech_parallel_wavegan.v1/checkpoint-400000steps.pkl",
                                  speed_control_alpha=1.0,)
wav = tts("Hi I am a good guy. This is a test.")["wav"]

write("x.wav", 22050, wav.view(-1).cpu().numpy())

while the same model works fine with old inference code.

text2speech = Text2Speech(
        train_config="exp/tts_train_raw_phn_tacotron_g2p_en_no_space/config.yaml",
        model_file="exp/tts_train_raw_phn_tacotron_g2p_en_no_space/200epoch.pth",
        device="cuda",
        # Only for Tacotron 2
        threshold=0.5,
        minlenratio=0.0,
        maxlenratio=10.0,
        use_att_constraint=False,
        backward_window=1,
        forward_window=3
    )

text2speech.spc2wav = None 

vocoder = load_model(config.vocoder_model).to("cuda").eval()
vocoder.remove_weight_norm()

with torch.no_grad():
    wav, outs, outs_denorm, probs, att_ws, durations, focus_rate = text2speech(text)
    wav = vocoder.inference(outs)

write("x.wav", 22050, wav.view(-1).cpu().numpy())

The new Text2Speech works perfectly fine with model_tag and vocoder_tag. I tried to use predefined model_tag (like kan-bayashi/ljspeech_tacotron2) with my trained vocoder (using new espent) and it still gives low volume outputs.

Basic environments:

• OS information: Ubuntu 18.04 x86_64]
• python version: new code python 3.8.12 and old code 3.8.7
• espnet version: above
• pytorch version: new code 1.10.2+cu102, old code 1.7.1

[kan-bayashi]

Great! Sometimes, we use different mean-var for text2mel and vocoder, so I decided to process as follows:

• text -> text2mel -> mel norm -> denorm by text2mel stats -> mel denorm -> norm by vocoder stats -> mel norm -> vocoder -> wav (with both stats case)
• text -> text2mel -> mel norm -> denorm by text2mel stats -> mel denorm -> vocoder -> wav (with only text2mel stats case)
• text -> text2mel -> mel -> vocoder -> wav (without stats case)

[kan-bayashi]

Thank you for your sharing. It seems normalization issue. Does /home/rohola/codes/ryan-tts/models/vocoders/train_nodev_ryanspeech_parallel_wavegan.v1 include stats.h5 file?