app.py

import torch
import soundfile as sf
import gradio as gr
from transformers import SpeechT5Processor, SpeechT5ForTextToSpeech
from transformers import SpeechT5HifiGan
from datasets import load_dataset
from IPython.display import Audio
import numpy as np
model_name = "trangiabao17032000/final_tts"

#processor
processor = SpeechT5Processor.from_pretrained(model_name)
tokenizer = processor.tokenizer

#model
model = SpeechT5ForTextToSpeech.from_pretrained(model_name)
model.resize_token_embeddings(len(tokenizer))
model.eval()

#vocoder
vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")
vocoder.eval()

#speaker embedding
embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)

#cleaner text
def convert_string_to_numbers(input_str):
    try:
        # Replace comma with period and attempt to convert the string to a float
        num = float(input_str.replace(',', '.'))
        if num.is_integer():
            return int(num)
        return num
    except ValueError:
        # If it's not a valid float, check if it's an integer or a negative integer
        if input_str.replace('.', '', 1).isdigit():  # Remove one dot for checking integers
            return int(input_str.replace(',', ''))
        elif input_str[0] == '-' and input_str[1:].replace('.', '', 1).isdigit():
            return int(input_str.replace(',', ''))
        else:
            raise ValueError("Invalid input: couldn't convert to a number")

def number_to_vietnamese_words(number):
    ones = ['', 'một', 'hai', 'ba', 'bốn', 'năm', 'sáu', 'bảy', 'tám', 'chín']
    tens = ['', 'mười', 'hai mươi', 'ba mươi', 'bốn mươi', 'năm mươi', 'sáu mươi', 'bảy mươi', 'tám mươi', 'chín mươi']
    hundreds = ['', 'một trăm', 'Hai trăm', 'ba trăm', 'bốn trăm', 'năm trăm', 'sáu trăm', 'bảy trăm', 'tám trăm', 'chín trăm']
    thousands = [''] + ['nghìn', 'triệu', 'tỷ']

    def words(n):

        if n < 10:
            return ones[n]
        elif n < 20:
            return tens[n//10]  + " " + words(n % 10)
        elif n < 100:
            return tens[n // 10] + ('' if n % 10 == 0 else ' ' + ones[n % 10])
        else:
            return hundreds[n // 100] + ('' if n % 100 == 0 else (' lẻ ' if n % 100 < 10 else ' ') + words(n % 100))

    if number == 0:
        return 'không'

    integer_part = int(number)
    decimal_part = round((number - integer_part) * 100)  # Round the decimal part to 2 decimal places

    result = []
    i = 0
    while integer_part > 0:
        if integer_part % 1000 != 0:
            result.append(words(integer_part % 1000) + (' ' + thousands[i] if i > 0 else ''))
        integer_part //= 1000
        i += 1

    result_integer = ' '.join(result[::-1])

    result_decimal = ''
    if decimal_part > 0:
        result_decimal = ' phẩy'
        for digit in str(decimal_part):
            result_decimal += ' ' + ones[int(digit)]

    return result_integer + result_decimal
    
def is_num(string):
    try:
        float(string)
    except ValueError:
        return False
    return True
    
def normalize(input):
     input = input.lower()
     newstr = map(lambda x: number_to_vietnamese_words(convert_string_to_numbers(x)) if is_num(x) else x, input.split(" "))
     return ' '.join(newstr)

def split_paragraph_into_sentences(paragraph, max_chars = 300):
    sentences = []
    words = paragraph.split()
    current_sentence = words[0]

    for word in words[1:]:
        if len(current_sentence) + len(word) + 1 <= max_chars:
            current_sentence += ' ' + word
        else:
            sentences.append(current_sentence)
            current_sentence = word

    if current_sentence:
        sentences.append(current_sentence)

    return sentences
    
# generator speech
def text_to_speech(paragraph):
    try:
        paragraph = normalize(paragraph)
    except:
        paragraph = paragraph.lower()
    list_sentence = split_paragraph_into_sentences(paragraph)
    final_speech = np.array([])
    
    for sentence in list_sentence:

      inputs = processor(text=sentence, return_tensors="pt")
      speech = model.generate_speech(inputs["input_ids"], speaker_embeddings,vocoder=vocoder)
      final_speech = np.concatenate((final_speech,  speech.numpy()))
      
    sf.write("tts_example.wav", final_speech, samplerate=16000)
    return "tts_example.wav"

tts_examples = [
    "xin chào mọi người, đây là sản phẩm thử nghiệm cho tiếng việt.",
    "Mình sẽ tổ chức sinh nhật vào thứ 6 ngày 7 tháng này",
]

#gradio interface
iface = gr.Interface(
    fn=text_to_speech,
    inputs=gr.Textbox(),
    outputs=gr.Audio(),
    title="Text-to-Speech",
    examples=tts_examples,
    description="Give me something to say!",
)

iface.launch()