app.py

import gradio as gr
import torch
from transformers import Wav2Vec2ForCTC, Wav2Vec2Tokenizer
import openai
import os

# --------------- STEP 1: Load the free STT model from Hugging Face ---------------
# We use a Persian Wav2Vec2 model
# Example: 'm3hrdadfi/wav2vec2-large-xlsr-persian-v2'
stt_model_name = "m3hrdadfi/wav2vec2-large-xlsr-persian-v2"
tokenizer = Wav2Vec2Tokenizer.from_pretrained(stt_model_name)
stt_model = Wav2Vec2ForCTC.from_pretrained(stt_model_name)

# --------------- STEP 2: Configure OpenAI API for GPT-4o mini (fine-tuned) ---------------
# You need your OpenAI API key. Let's read from environment variable for safety.
openai.api_key = os.getenv("OPENAI_API_KEY", "YOUR_API_KEY_HERE") 
# If you'd rather hardcode, replace "YOUR_API_KEY_HERE" with your actual key. 
# But do not commit your real key publicly.

OPENAI_MODEL_NAME = "gpt-4o-mini"  # The exact model name you said you'd use (example)

# --------------- STT Function ---------------
def speech_to_text(audio):
    """
    audio is a tuple (sample_rate, numpy_data) from Gradio's microphone or an audio file.
    Convert Persian speech to text using wav2vec2 model offline.
    """
    if audio is None:
        return ""
    sample_rate, data = audio
    # Convert to torch tensor
    input_values = tokenizer(data, return_tensors="pt", sampling_rate=sample_rate).input_values
    with torch.no_grad():
        logits = stt_model(input_values).logits
    predicted_ids = torch.argmax(logits, dim=-1)
    transcription = tokenizer.decode(predicted_ids[0])
    # transcription might be uppercase and missing punctuation. We just return it raw.
    return transcription.lower()

# --------------- Correction Function with GPT-4o mini ---------------
def correct_text_with_gpt(text):
    """
    Send the text to OpenAI GPT-4o mini for correction/improvement.
    The user wants a formal tone consistent with medical/legal usage.
    Fine-tuned model is presumably specialized for that domain.
    """
    if not text.strip():
        return ""
    
    system_message = (
        "You are a specialized model that receives Persian text and corrects it with a formal tone, "
        "particularly for medical/legal context. Make the text coherent, unify spacing, and punctuation, "
        "and ensure a formal writing style. Do not add commentary. Just provide the corrected text."
    )
    
    user_message = f"متن خام: {text}\n\nلطفاً متن فوق را در یک پاراگراف رسمی و اداری اصلاح کن."

    # We'll call the ChatCompletion API if GPT-4o mini is chat-based.
    # If it's a completion-based model, code differs. But let's assume it's chat-based for simplicity.
    response = openai.ChatCompletion.create(
        model=OPENAI_MODEL_NAME,
        messages=[
            {"role": "system", "content": system_message},
            {"role": "user", "content": user_message}
        ],
        temperature=0.2,
        max_tokens=1024
    )
    corrected = response.choices[0].message.content
    return corrected.strip()

# --------------- Gradio Interface ---------------
# We want:
# - A widget to record or upload audio
# - A box that shows transcribed text
# - A button "اصلاح متن" that calls GPT-4o mini
# - A final output box

def process_audio_and_correct(audio):
    # Step A: convert speech to text
    raw_text = speech_to_text(audio)
    # Step B: correct with GPT
    final_text = correct_text_with_gpt(raw_text)
    return raw_text, final_text

with gr.Blocks() as demo:
    gr.Markdown("# وب‌سایت تبدیل گفتار به متن (فارسی) + اصلاح متن با GPT-4o mini")
    gr.Markdown("در اینجا می‌توانید یک وویس فارسی ضبط یا فایل صوتی آپلود کنید. سپس متن استخراج‌شده و اصلاح‌شده را دریافت کنید.")
    
    with gr.Row():
        audio_in = gr.Audio(source="microphone", type="numpy", label="ضبط صدا یا آپلود فایل (رایگان)")
    
    transcribed_text = gr.Textbox(label="متن خام (خروجی مرحله اول - رایگان)", lines=3)
    corrected_text = gr.Textbox(label="متن اصلاح‌شده (با GPT-4o mini - پولی)", lines=3)
    
    btn_process = gr.Button("تبدیل و اصلاح")
    
    def audio_workflow(audio):
        # step 1: stt
        raw_text = speech_to_text(audio)
        transcribed_text.update(raw_text)
        # step 2: correction
        corrected = correct_text_with_gpt(raw_text)
        corrected_text.update(corrected)
        return raw_text, corrected
    
    btn_process.click(fn=process_audio_and_correct, inputs=audio_in, outputs=[transcribed_text, corrected_text])

demo.launch()