app.py

import gradio as gr
from transformers import pipeline, MarianMTModel, MarianTokenizer
import edge_tts
import asyncio

# Load Whisper for English ASR
asr_en = pipeline("automatic-speech-recognition", model="openai/whisper-base")

# Load MarianMT for Yoruba-English (can switch for Igbo/Hausa)
mt_model_name = "Helsinki-NLP/opus-mt-yo-en"
tokenizer = MarianTokenizer.from_pretrained(mt_model_name)
model = MarianMTModel.from_pretrained(mt_model_name)

def translate_speech(audio):
    transcription = asr_en(audio)["text"]
    inputs = tokenizer(transcription, return_tensors="pt", padding=True)
    translated_tokens = model.generate(**inputs)
    translated_text = tokenizer.decode(translated_tokens[0], skip_special_tokens=True)
    return transcription, translated_text

def translate_text(text):
    inputs = tokenizer(text, return_tensors="pt", padding=True)
    translated_tokens = model.generate(**inputs)
    translated_text = tokenizer.decode(translated_tokens[0], skip_special_tokens=True)
    return translated_text

with gr.Blocks() as demo:
    gr.Markdown("### Multilingual Realtime Translator - English ↔ Yoruba (V1)")
    
    with gr.Tab("Speech Input"):
        mic = gr.Audio(source="microphone", type="filepath", label="Speak Now")
        transcription = gr.Textbox(label="Transcription")
        translation = gr.Textbox(label="Translation")
        mic.submit(translate_speech, inputs=mic, outputs=[transcription, translation])

    with gr.Tab("Text Input"):
        input_text = gr.Textbox(label="Enter text")
        translated_text = gr.Textbox(label="Translated text")
        input_text.submit(translate_text, inputs=input_text, outputs=translated_text)

demo.launch()