import os
import gradio as gr
import whisper
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
from gtts import gTTS

def translate_voice(file, target_lang):
    # Auto to text (STT) 
    model = whisper.load_model("base")
    audio = whisper.load_audio(file.name)
    audio = whisper.pad_or_trim(audio)
    mel = whisper.log_mel_spectrogram(audio).to(model.device).float() 
    _, probs = model.detect_language(mel)

    options = whisper.DecodingOptions()
    result = whisper.decode(model, mel, options)

    text = result.text
    lang = max(probs, key=probs.get)

    # Translate
    tokenizer = AutoTokenizer.from_pretrained("alirezamsh/small100")
    model = AutoModelForSeq2SeqLM.from_pretrained("alirezamsh/small100")

    tokenizer.src_lang = target_lang
    encoded_bg = tokenizer(text, return_tensors="pt")
    generated_tokens = model.generate(**encoded_bg)
    translated_text = tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)[0]

    # Text-to-audio (TTS)
    tts = gTTS(text=translated_text, lang=target_lang)
    filename = "to_speech.mp3"
    tts.save(filename)

    return filename, text, translated_text, target_lang

iface = gr.Interface(
    fn=translate_voice, 
    inputs=[
        gr.components.File(type="file", label="Your Audio"), 
        gr.components.Dropdown(choices=['en', 'ru', 'de', 'fr'], label="Target Language")
    ], 
    outputs=[
        gr.components.Audio(type="filepath", label="Translated Audio"),
        gr.components.Textbox(label="Original Text"),
        gr.components.Textbox(label="Translated Text"),
        gr.components.Textbox(label="Target Language"),
    ]
)
iface.launch()