app.py

import os
import subprocess
from os.path import exists, join, expanduser
from huggingface_hub import snapshot_download

# project_name = "pytorch-dc-tts"
# if not exists(project_name):
#   ! git clone --quiet https://github.com/tugstugi/{project_name}
#   ! cd {project_name} && pip install -q -r requirements.txt

project_name = "pytorch-dc-tts"
if not os.path.exists(project_name):
    subprocess.run(["git", "clone", "--quiet", f"https://github.com/tugstugi/{project_name}"])

import os
import shutil

project_name = "pytorch-dc-tts"

# Get a list of all files and folders in the project directory
for item in os.listdir(project_name):
    source = os.path.join(project_name, item)
    destination = os.path.join(".", item)  # Move to the current (main) directory

    # Move the item, handling files and folders differently
    if os.path.isfile(source):
        shutil.move(source, destination)
    elif os.path.isdir(source):
        shutil.move(source, destination)

# (Optional) Remove the original project directory if it's empty
if not os.listdir(project_name):
    os.rmdir(project_name)

#download model
snapshot_download(repo_id="Dorjzodovsuren/Mongolian_TTS", local_dir="./")

import sys
import uuid
import torch
import gradio as gr
#sys.path.append(project_name)

import warnings
warnings.filterwarnings("ignore")  # ignore warnings in this notebook

import numpy as np
from tqdm import *

from hparams import HParams as hp
from audio import save_to_wav
from models import Text2Mel, SSRN
from datasets.mb_speech import vocab, idx2char, get_test_data, number2word

hp.max_T = 300

# Check for CUDA availability and set device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")


def load_checkpoint(checkpoint_file_name, model, optimizer):
    """Loads the checkpoint into the given model and optimizer."""
    checkpoint = torch.load(checkpoint_file_name, map_location=torch.device(device))
    model.load_state_dict(checkpoint['state_dict'])
    model.float()
    if optimizer is not None:
        optimizer.load_state_dict(checkpoint['optimizer'])
    start_epoch = checkpoint.get('epoch', 0)
    global_step = checkpoint.get('global_step', 0)
    del checkpoint
    print("loaded checkpoint epoch=%d step=%d" % (start_epoch, global_step))
    return start_epoch, global_step


text2mel = Text2Mel(vocab).eval()
last_checkpoint_file_name = "text2mel.pth"
if last_checkpoint_file_name:
    print("loading text2mel checkpoint '%s'..." % last_checkpoint_file_name)
    load_checkpoint(last_checkpoint_file_name, text2mel, None)


ssrn = SSRN().eval()
last_checkpoint_file_name = "ssrn.pth"
if last_checkpoint_file_name:
    print("loading text2mel checkpoint '%s'..." % last_checkpoint_file_name)
    load_checkpoint(last_checkpoint_file_name, ssrn, None)

def generate_random_id():
  """Generates a random UUID (Universally Unique Identifier).

  Returns:
    A random UUID as a string.
  """
  return str(uuid.uuid4())

# synthetize by one by one because there is a batch processing bug!
def text_to_speech(sentence_raw):
    sentence = ' '.join([number2word(s) if s.isdigit() else s for s in sentence_raw.split()])
    normalized_sentence = "".join([c if c.lower() in vocab else '' for c in sentence])
    print(normalized_sentence)

    sentences = [normalized_sentence]
    max_N = len(normalized_sentence)
    L = torch.from_numpy(get_test_data(sentences, max_N))
    zeros = torch.from_numpy(np.zeros((1, hp.n_mels, 1), np.float32))
    Y = zeros
    A = None

    for t in range(hp.max_T):
      _, Y_t, A = text2mel(L, Y, monotonic_attention=True)
      Y = torch.cat((zeros, Y_t), -1)
      _, attention = torch.max(A[0, :, -1], 0)
      attention = attention.item()
      if L[0, attention] == vocab.index('E'):  # EOS
          break

    _, Z = ssrn(Y)

    Z = Z.cpu().detach().numpy()
    random_id = generate_random_id()
    save_to_wav(Z[0, :, :].T, f'{random_id}.wav')

    return f"{random_id}.wav"

# Gradio interface
iface = gr.Interface(
    fn=text_to_speech,
    inputs="text",
    outputs="audio",
    title="Mongolian Text to Speech",
    description="Энэхүү Монгол хэлний текстээс яриа үүсгэх системийг Hugging Face дээр ашиглах боломжтой боллоо. Монгол хэлээр бичсэн текстийг оруулснаар яриа хэлбэрээр буцаан авах боломжтой бөгөөд үүнийг та үнэ төлбөргүйгээр хүссэн үедээ ашиглах боломжтой . Одоогоор зөвхөн CPU дээр ажиллаж байгаа тул боловсруулалт удаан байж магадгүй. Enjoy ;D",
    examples=[["Сайн байна уу, найзууд аа."], ["Монгол улсын нийслэл Улаанбаатараас ярьж байна."], ["Чихинү чимэг болсон аялгуу сайхан монгол хэл, Чин зоригт өвгөд дээдсийн минь өв их эрдэнэ."]],
    cache_examples=True
)

iface.queue().launch()