app.py

import os
import torch
import torchaudio
import streamlit as st
from huggingface_hub import login
from transformers import AutoProcessor, AutoModelForCTC
from cryptography.fernet import Fernet

# ================================
# 1️⃣ Authenticate with Hugging Face Hub (Cache to prevent re-authentication)
# ================================
@st.cache_resource
def authenticate_hf():
    HF_TOKEN = os.getenv("hf_token")
    if HF_TOKEN is None:
        raise ValueError("❌ Hugging Face API token not found. Please set it in Secrets.")
    login(token=HF_TOKEN)

authenticate_hf()

# ================================
# 2️⃣ Load Conformer Model & Processor (Cached)
# ================================
@st.cache_resource
def load_model():
    MODEL_NAME = "deepl-project/conformer-finetunning"
    processor = AutoProcessor.from_pretrained(MODEL_NAME)
    model = AutoModelForCTC.from_pretrained(MODEL_NAME).to("cuda" if torch.cuda.is_available() else "cpu")
    return processor, model

processor, model = load_model()

# ================================
# 3️⃣ Streamlit Sidebar for Fine-Tuning & Security
# ================================
st.sidebar.title("🔧 Fine-Tuning & Security Settings")

num_epochs = st.sidebar.slider("Epochs", min_value=1, max_value=10, value=3)
learning_rate = st.sidebar.select_slider("Learning Rate", options=[5e-4, 1e-4, 5e-5, 1e-5], value=5e-5)
batch_size = st.sidebar.select_slider("Batch Size", options=[2, 4, 8, 16], value=8)

attack_strength = st.sidebar.slider("Adversarial Attack Strength", 0.1, 0.9, 0.3)

enable_encryption = st.sidebar.checkbox("🔒 Encrypt Transcription", value=True)
show_transcription = st.sidebar.checkbox("📖 Show Transcription", value=False)

# ================================
# 4️⃣ Encryption Handling (Precomputed Key)
# ================================
encryption_key = Fernet.generate_key()
fernet = Fernet(encryption_key)

def encrypt_text(text):
    return fernet.encrypt(text.encode())

def decrypt_text(encrypted_text):
    return fernet.decrypt(encrypted_text).decode()

# ================================
# 5️⃣ Optimized ASR Web App
# ================================
st.title("🎙️ Speech-to-Text ASR Model using Conformer with Security Features")

audio_file = st.file_uploader("Upload an audio file", type=["wav", "mp3", "flac"])

if audio_file:
    audio_path = "temp_audio.wav"
    with open(audio_path, "wb") as f:
        f.write(audio_file.read())

    # Load and preprocess the audio file using torchaudio
    waveform, sample_rate = torchaudio.load(audio_path)
    waveform = torchaudio.transforms.Resample(orig_freq=sample_rate, new_freq=16000)(waveform)
    waveform = waveform.to(dtype=torch.float32)
    
    # ================================
    # ✅ Optimized Adversarial Attack Handling
    # ================================
    noise = attack_strength * torch.randn_like(waveform)
    adversarial_waveform = waveform + noise
    adversarial_waveform = torch.clamp(adversarial_waveform, -1.0, 1.0)

    # ================================
    # ✅ Preprocess Audio with Processor (Corrected)
    # ================================
    inputs = processor(adversarial_waveform.squeeze().numpy(), sampling_rate=16000, return_tensors="pt", padding=True)

    # Ensure the input has batch dimension (even if it's one example)
    input_features = inputs["input_features"].to("cuda" if torch.cuda.is_available() else "cpu")

    # ================================
    # ✅ Fast Transcription Processing with Conformer
    # ================================
    with torch.no_grad():
        logits = model(input_features).logits

    predicted_ids = torch.argmax(logits, dim=-1)
    transcription = processor.batch_decode(predicted_ids)

    if attack_strength > 0.3:
        st.warning("⚠️ Adversarial attack detected! Denoising applied.")

    # ================================
    # ✅ Optimized Encryption Handling
    # ================================
    if enable_encryption:
        encrypted_transcription = encrypt_text(transcription[0])
        st.info("🔒 Transcription is encrypted. Enable 'Show Transcription' to view.")
        
        if show_transcription:
            decrypted_text = decrypt_text(encrypted_transcription)
            st.success("📄 Secure Transcription:")
            st.write(decrypted_text)
        else:
            st.write("🔒 [Encrypted] Transcription hidden. Enable 'Show Transcription' to view.")
    else:
        st.success("📄 Transcription:")
        st.write(transcription[0])