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

model_id = "ovieyra21/speecht5_tts_mabama_nl"  # 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 = [
    ("à", "a"),
    ("â", "a"),
    ("ç", "c"),
    ("è", "e"),
    ("ë", "e"),
    ("î", "i"),
    ("ï", "i"),
    ("ô", "o"),
    ("ù", "u"),
    ("û", "u"),
    ("ü", "u"),
]

title = "Text-to-Speech"
description = """
Demo for text-to-speech translation in French. Demo uses [Sandiago21/speecht5_finetuned_facebook_voxpopuli_french](https://huggingface.co/Sandiago21/speecht5_finetuned_facebook_voxpopuli_french) checkpoint, which is based on Microsoft's
[SpeechT5 TTS](https://huggingface.co/microsoft/speecht5_tts) model and is fine-tuned in French 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 (16000, speech.cpu().numpy())  # Devuelve el audio directamente

syntesize_speech_gradio = gr.Interface(
    synthesize_speech,
    inputs=gr.Textbox(label="Text", placeholder="Type something here..."),
    outputs=gr.Audio(),
    examples=["Probando audio"],
    title=title,
    description=description,
).launch()