app.py

import gradio as gr
import torch
from datasets import load_dataset
from transformers import pipeline, SpeechT5Processor, SpeechT5HifiGan, SpeechT5ForTextToSpeech

model_id = "Sandiago21/speecht5_finetuned_mozilla_foundation_common_voice_13_german"  # update with your model id
model = SpeechT5ForTextToSpeech.from_pretrained(model_id)
vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")
embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
speaker_embeddings = torch.tensor(embeddings_dataset[7440]["xvector"]).unsqueeze(0)

processor = SpeechT5Processor.from_pretrained(model_id)

replacements = [    
    ("Ä", "E"),
    ("Æ", "E"),
    ("Ç", "C"),
    ("É", "E"),
    ("Í", "I"),
    ("Ó", "O"),
    ("Ö", "E"),
    ("Ü", "Y"),
    ("ß", "S"),
    ("à", "a"),
    ("á", "a"),
    ("ã", "a"),
    ("ä", "e"),
    ("å", "a"),
    ("ë", "e"),
    ("í", "i"),
    ("ï", "i"),
    ("ð", "o"),
    ("ñ", "n"),
    ("ò", "o"),
    ("ó", "o"),
    ("ô", "o"),
    ("ö", "u"),
    ("ú", "u"),
    ("ü", "y"),
    ("ý", "y"),
    ("Ā", "A"),
    ("ā", "a"),
    ("ă", "a"),
    ("ą", "a"),
    ("ć", "c"),
    ("Č", "C"),
    ("č", "c"),
    ("ď", "d"),
    ("Đ", "D"),
    ("ę", "e"),
    ("ě", "e"),
    ("ğ", "g"),
    ("İ", "I"),
    ("О", "O"),
    ("Ł", "L"),
    ("ń", "n"),
    ("ň", "n"),
    ("Ō", "O"),
    ("ō", "o"),
    ("ő", "o"),
    ("ř", "r"),
    ("Ś", "S"),
    ("ś", "s"),
    ("Ş", "S"),
    ("ş", "s"),
    ("Š", "S"),
    ("š", "s"),
    ("ū", "u"),
    ("ź", "z"),
    ("Ż", "Z"),
    ("Ž", "Z"),
    ("ǐ", "i"),
    ("ǐ", "i"),
    ("ș", "s"),
    ("ț", "t"),
]


title = "Text-to-Speech"
description = """
Demo for text-to-speech translation in German. Demo uses [Sandiago21/speecht5_finetuned_mozilla_foundation_common_voice_13_german](https://huggingface.co./Sandiago21/speecht5_finetuned_mozilla_foundation_common_voice_13_german) checkpoint, which is based on Microsoft's
[SpeechT5 TTS](https://huggingface.co./microsoft/speecht5_tts) model and is fine-tuned in German Audio dataset
![Text-to-Speech (TTS)"](https://geekflare.com/wp-content/uploads/2021/07/texttospeech-1200x385.png "Diagram of Text-to-Speech (TTS)")
"""


def cleanup_text(text):
    for src, dst in replacements:
        text = text.replace(src, dst)
    return text

def synthesize_speech(text):
    text = cleanup_text(text)
    inputs = processor(text=text, return_tensors="pt")

    speech = model.generate_speech(inputs["input_ids"], speaker_embeddings, vocoder=vocoder)

    return gr.Audio.update(value=(16000, speech.cpu().numpy()))

syntesize_speech_gradio = gr.Interface(
    synthesize_speech,
    inputs = gr.Textbox(label="Text", placeholder="Type something here..."),
    outputs=gr.Audio(),
    examples=["Daher wird die Reform der Europäischen Sozialfondsverordnung, die wir morgen beschließen, auch umgehend in Kraft treten."],
    title=title,
    description=description,
).launch()