app.py

import gradio as gr

from tacotron2_gst_original.hyper_parameters import tacotron_params as hparams
from tacotron2_gst_original.training import load_model

from tacotron2_gst_original.text import text_to_sequence

from melgan.model.generator import Generator
from melgan.utils.hparams import load_hparam

torch.manual_seed(1234)
MAX_WAV_VALUE = 32768.0

def init_models(hparams):
    # load trained tacotron2 + GST model:
    model = load_model(hparams)
    checkpoint_path = "trained_models/checkpoint_78000.model"
    model.load_state_dict(torch.load(checkpoint_path)['state_dict'])
    model.to('cuda')
    _ = model.eval()
    
    # load pre trained MelGAN model for mel2audio:
    vocoder_checkpoint_path = "trained_models/nvidia_tacotron2_LJ11_epoch6400.pt"
    checkpoint = torch.load(vocoder_checkpoint_path)
    hp_melgan = load_hparam("melgan/config/default.yaml")
    vocoder_model = Generator(80)
    vocoder_model.load_state_dict(checkpoint['model_g'])
    vocoder_model = vocoder_model.to('cuda')
    vocoder_model.eval(inference=False)

def synthesize(text):
    sequence = np.array(text_to_sequence(text, ['english_cleaners']))[None, :]
    sequence = torch.from_numpy(sequence).to(device='cuda', dtype=torch.int64)

    gst_head_scores = np.array([0.5, 0.15, 0.35])  # originally ([0.5, 0.15, 0.35])
    gst_scores = torch.from_numpy(gst_head_scores).cuda().float()

    mel_outputs, mel_outputs_postnet, _, alignments = model.inference(sequence, gst_scores)

    # mel2wav inference:
    with torch.no_grad():
      audio = vocoder_model.inference(mel_outputs_postnet)
    
    audio_numpy = audio.data.cpu().detach().numpy()

    return (22050, audio_numpy)


init_models(hparams)
iface = gr.Interface(fn=synthesize, inputs="text", outputs=[gr.Audio(label="Generated Speech", type="numpy"),])
iface.launch()