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

device = "cuda" if torch.cuda.is_available() else "cpu"

def load_models_and_data():
    model_name = "microsoft/speecht5_tts"
    processor = SpeechT5Processor.from_pretrained(model_name)
    model = SpeechT5ForTextToSpeech.from_pretrained("emirhanbilgic/speecht5_finetuned_emirhan_tr").to(device)
    vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)
    
    embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
    speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0).to(device)
    
    return model, processor, vocoder, speaker_embeddings

model, processor, vocoder, speaker_embeddings = load_models_and_data()

@spaces.GPU(duration = 60)
def text_to_speech(text):
    inputs = processor(text=text, return_tensors="pt").to(device)
    speech = model.generate_speech(inputs["input_ids"], speaker_embeddings, vocoder=vocoder)
    sf.write("output.wav", speech.cpu().numpy(), samplerate=16000)
    return "output.wav"

iface = gr.Interface(
    fn=text_to_speech,
    inputs=gr.Textbox(label="Enter Turkish text to convert to speech"),
    outputs=gr.Audio(label="Generated Speech"),
    title="Turkish SpeechT5 Text-to-Speech Demo",
    description="Enter Turkish text and listen to the generated speech using the fine-tuned SpeechT5 model."
)

iface.launch()