app.py

# app_gradio.py
import gradio as gr
import numpy as np
import torch
import soundfile as sf
import os
import yaml
from dotenv import load_dotenv
from threading import Thread

# --- TTS & AI Imports ---
from parler_tts import ParlerTTSForConditionalGeneration
from transformers import AutoTokenizer, AutoFeatureExtractor, set_seed
from streamer import ParlerTTSStreamer # Make sure streamer.py is available

from src.detection.factory import get_detector
from src.alerting.alert_system import get_alerter

# --- Load Configuration and Environment Variables ---
# This part is the same as our Streamlit app
load_dotenv()
config_path = 'config.yaml'
with open(config_path, 'r') as f:
    config = yaml.safe_load(f)
secrets = {
    "gemini_api_key": os.getenv("GEMINI_API_KEY"),
}

# --- Initialize Backend Components ---
print("Initializing detector and alerter...")
detector = get_detector(config)
alerter = get_alerter(config, secrets["gemini_api_key"])
print("Initialization complete. Launching UI...")

# --- Parler-TTS Model Setup (Requires GPU) ---
print("Loading Parler-TTS model. This may take a moment...")
device = "cuda:0" if torch.cuda.is_available() else "cpu"
if device == "cpu":
    print("\nWARNING: Running Parler-TTS on a CPU will be extremely slow. A GPU is highly recommended.\n")
torch_dtype = torch.float16 if device != "cpu" else torch.float32


# Using a smaller, faster model suitable for real-time alerts
repo_id = "parler-tts/parler_tts_mini_v0.1"
model = ParlerTTSForConditionalGeneration.from_pretrained(repo_id, torch_dtype=torch_dtype).to(device)
tokenizer = AutoTokenizer.from_pretrained(repo_id)
feature_extractor = AutoFeatureExtractor.from_pretrained(repo_id)
print("Parler-TTS model loaded.")

# --- Audio Streaming Generator Function ---
def stream_alert_audio(text_prompt):
    """
    A generator function that yields audio chunks for a given text prompt.
    This is the core of the streaming implementation.
    """
    sampling_rate = model.config.sampling_rate
    description = "Jenny is A female speaker with a clear and urgent voice." # Voice prompt for TTS
    
    prompt_ids = tokenizer(text_prompt, return_tensors="pt").input_ids.to(device)
    description_ids = tokenizer(description, return_tensors="pt").input_ids.to(device)

    # Setup the streamer
    streamer = ParlerTTSStreamer(model, device, play_steps=int(sampling_rate * 2.0))
    
    generation_kwargs = dict(
        input_ids=description_ids,
        prompt_input_ids=prompt_ids,
        streamer=streamer,
        do_sample=True,
        temperature=1.0, # Increase for more vocal variety
        repetition_penalty=1.2,
    )
    
    # Run generation in a separate thread to not block the UI
    thread = Thread(target=model.generate, kwargs=generation_kwargs)
    
    try:
        thread.start()
        print(f"Audio stream started for: '{text_prompt}'")
        # Yield audio chunks as they become available
        for new_audio_chunk in streamer:
            yield (sampling_rate, new_audio_chunk)
    finally:
        # CRITICAL: This block runs after the generator is exhausted (audio finishes)
        # We reset the alerter state so that a new alert can be triggered later.
        print("Audio stream finished. Resetting alerter state.")
        alerter.reset_alert()
        
# --- Main Webcam Processing Function ---
def process_live_frame(frame):
    """
    Processes each webcam frame, performs drowsiness detection, and
    returns a generator for audio streaming when an alert is triggered.
    """
    if frame is None:
        return np.zeros((480, 640, 3), dtype=np.uint8), "Status: Inactive", None

    processed_frame, indicators, _ = detector.process_frame(frame)
    drowsiness_level = indicators.get("drowsiness_level", "Awake")
    lighting = indicators.get("lighting", "Good")
    score = indicators.get("details", {}).get("Score", 0)

    # Build status text
    status_text = f"Lighting: {lighting}\n"
    if lighting == "Low":
        status_text += "Detection paused due to low light."
    else:
        status_text += f"Status: {drowsiness_level}\nScore: {score:.2f}"

    # --- Alert Trigger Logic ---
    audio_output = None
    if drowsiness_level != "Awake":
        # alerter.trigger_alert() returns the alert TEXT if not on cooldown, otherwise None.
        alert_text = alerter.trigger_alert(level=drowsiness_level)
        if alert_text:
            # If we got text, it means we can start an alert.
            # We return the generator function itself. Gradio will handle it.
            audio_output = stream_alert_audio(alert_text)
            
    # On subsequent frames where the user is drowsy, trigger_alert() will return None
    # due to the cooldown, preventing a new stream from starting, which is what we want.

    return processed_frame, status_text, audio_output


# --- Gradio UI Definition ---
with gr.Blocks(theme=gr.themes.Default(primary_hue="blue")) as app:
    gr.Markdown("# 🚗 Drive Paddy - Drowsiness Detection (Streaming)")
    gr.Markdown("Live drowsiness detection with real-time, streaming voice alerts.")

    with gr.Row():
        with gr.Column(scale=2):
            webcam_input = gr.Image(sources=["webcam"], streaming=True, label="Live Camera Feed")
        with gr.Column(scale=1):
            processed_output = gr.Image(label="Processed Feed")
            status_output = gr.Textbox(label="Live Status", lines=3, interactive=False)
            
            # --- KEY CHANGE: The Audio component now uses streaming=True ---
            audio_alert_output = gr.Audio(
                label="Alert System", 
                autoplay=True, 
                visible=False, # Hide the player controls
                streaming=True
            )

    webcam_input.stream(
        fn=process_live_frame,
        inputs=[webcam_input],
        outputs=[processed_output, status_output, audio_alert_output]
    )


# --- Launch the App ---
if __name__ == "__main__":
    app.launch(debug=True)