app.py

import gradio as gr
import time
import urllib.request
from pathlib import Path
import os
import torch
import numpy
import scipy.io.wavfile
from espnet2.bin.tts_inference import Text2Speech
from espnet2.utils.types import str_or_none


# def load_model(model_tag, vocoder_tag):
#     from espnet_model_zoo.downloader import ModelDownloader

#     kwargs = {}

#     # Model
#     d = ModelDownloader()
#     kwargs = d.download_and_unpack(model_tag)

#     # Vocoder
#     download_dir = Path(os.path.expanduser("~/.cache/parallel_wavegan"))
#     vocoder_dir = download_dir / vocoder_tag
#     os.makedirs(vocoder_dir, exist_ok=True)

#     kwargs["vocoder_config"] = vocoder_dir / "config.yml"
#     if not kwargs["vocoder_config"].exists():
#         urllib.request.urlretrieve(f"https://huggingface.co/{vocoder_tag}/resolve/main/config.yml", kwargs["vocoder_config"])

#     kwargs["vocoder_file"] = vocoder_dir / "checkpoint-50000steps.pkl"
#     if not kwargs["vocoder_file"].exists():
#         urllib.request.urlretrieve(f"https://huggingface.co/{vocoder_tag}/resolve/main/checkpoint-50000steps.pkl", kwargs["vocoder_file"])

#     return Text2Speech(
#         **kwargs,
#         device="cpu",
#         threshold=0.5,
#         minlenratio=0.0,
#         maxlenratio=10.0,
#         use_att_constraint=True,
#         backward_window=1,
#         forward_window=4,
#     )

# gos_text2speech = load_model('https://huggingface.co/wietsedv/tacotron2-gronings/resolve/main/tts_ljspeech_finetune_tacotron2.v5_train.loss.ave.zip', 'wietsedv/parallelwavegan-gronings')
# nld_text2speech = load_model('https://huggingface.co/wietsedv/tacotron2-dutch/resolve/main/tts_ljspeech_finetune_tacotron2.v5_train.loss.ave.zip', 'wietsedv/parallelwavegan-dutch')

gos_text2speech = Text2Speech.from_pretrained(
   model_tag="https://huggingface.co/wietsedv/tacotron2-gronings/resolve/main/tts_ljspeech_finetune_tacotron2.v5_train.loss.ave.zip",
   vocoder_tag="parallel_wavegan/ljspeech_parallel_wavegan.v3",
   device="cpu",
   threshold=0.5,
   minlenratio=0.0,
   maxlenratio=10.0,
   use_att_constraint=True,
   backward_window=1,
   forward_window=4,
)
nld_text2speech = Text2Speech.from_pretrained(
   model_tag="https://huggingface.co/wietsedv/tacotron2-dutch/resolve/main/tts_ljspeech_finetune_tacotron2.v5_train.loss.ave.zip",
   vocoder_tag="parallel_wavegan/ljspeech_parallel_wavegan.v3",
   device="cpu",
   threshold=0.5,
   minlenratio=0.0,
   maxlenratio=10.0,
   use_att_constraint=True,
   backward_window=1,
   forward_window=4,
)
#eng_text2speech = Text2Speech.from_pretrained(
#    model_tag="kan-bayashi/ljspeech_tacotron2",
#    vocoder_tag="parallel_wavegan/ljspeech_parallel_wavegan.v3",
#    device="cpu",
#    threshold=0.5,
#    minlenratio=0.0,
#    maxlenratio=10.0,
#    use_att_constraint=True,
#    backward_window=1,
#    forward_window=4,
#)

def inference(text,lang):
  with torch.no_grad():
      if lang == "gronings":
          wav = gos_text2speech(text)["wav"]
          scipy.io.wavfile.write("out.wav", gos_text2speech.fs , wav.view(-1).cpu().numpy())
      if lang == "dutch":
          wav = nld_text2speech(text)["wav"]
          scipy.io.wavfile.write("out.wav", nld_text2speech.fs , wav.view(-1).cpu().numpy())
      #if lang == "english":
      #    wav = eng_text2speech(text)["wav"]
      #    scipy.io.wavfile.write("out.wav", eng_text2speech.fs , wav.view(-1).cpu().numpy())

  return  "out.wav", "out.wav"

title = "GroTTS"
examples = [
  ['Ze gingen mit klas noar waddendiek, over en deur bragel lopen.', 'gronings']
]

gr.Interface(
    inference,
    [gr.inputs.Textbox(label="input text", lines=3), gr.inputs.Radio(choices=["gronings", "dutch"], type="value", default="gronings", label="language")], 
    [gr.outputs.Audio(type="file", label="Output"), gr.outputs.File()],
    title=title,
    examples=examples
    ).launch(enable_queue=True)