app.py

import gradio as gr
import numpy as np
import librosa
import requests
from io import BytesIO
from PIL import Image
import os
from tensorflow.keras.models import load_model
from faster_whisper import WhisperModel

# Load the emotion prediction model
def load_emotion_model(model_path):
    try:
        model = load_model(model_path)
        return model
    except Exception as e:
        print("Error loading emotion prediction model:", e)
        return None

model_path = 'mymodel_SER_LSTM_RAVDESS.h5'
model = load_emotion_model(model_path)

# Initialize WhisperModel
model_size = "small"
model2 = WhisperModel(model_size, device="cpu", compute_type="int8")

# Function to transcribe audio
def transcribe(wav_filepath):
    segments, _ = model2.transcribe(wav_filepath, beam_size=5)
    return "".join([segment.text for segment in segments])

# Function to extract MFCC features from audio
def extract_mfcc(wav_file_name):
    try:
        y, sr = librosa.load(wav_file_name)
        mfccs = np.mean(librosa.feature.mfcc(y=y, sr=sr, n_mfcc=40).T, axis=0)
        return mfccs
    except Exception as e:
        print("Error extracting MFCC features:", e)
        return None

# Emotions dictionary (aligned with training: 0–7)
emotions = {
    0: 'neutral',
    1: 'calm',
    2: 'happy',
    3: 'sad',
    4: 'angry',
    5: 'fearful',
    6: 'disgust',
    7: 'surprised'
}

# Function to predict emotion from audio
def predict_emotion_from_audio(wav_filepath):
    try:
        test_point = extract_mfcc(wav_filepath)

        if test_point is None:
            return "Error: Unable to extract features"

        # Reshape to match model input
        test_point = np.reshape(test_point, (1, 40, 1))

        # Predict
        predictions = model.predict(test_point, verbose=0)

        # Get index (0–7, consistent with training)
        predicted_emotion_label = np.argmax(predictions[0])

        return emotions[predicted_emotion_label]

    except Exception as e:
        print("Error predicting emotion:", e)
        return None


api_key = os.getenv("DeepAI_api_key")

# Function to generate an image using DeepAI Text to Image API


import random

def generate_image(emotion_prediction, transcribed_text, output_resolution=(1024, 1024)):

    try:
        url = "https://api.deepai.org/api/image-editor"
        headers = {
            'api-key': api_key
        }
        # Select a random image file from TerraIncognita0.jpg to TerraIncognita9.jpg
        image_file_path = f'TAI_Images/TerraIncognita{random.randint(0, 9)}.jpg'
        files = {
            'image': open(image_file_path, 'rb'),
            'text': f"""
        Transform this image into an ancient exploratory map of an unknown land. The territory emerges from: {transcribed_text}. Its geography, symbols, and distortions should evoke the emotion of {emotion_prediction}."""
        }
        response = requests.post(url, headers=headers, files=files)
        response_data = response.json()
        if 'output_url' in response_data:
            return response_data['output_url']
        else:
            return None
    except Exception as e:
        print("Error generating image:", e)
        return None

        
# Function to get predictions
def get_predictions(audio_input):
    emotion_prediction = predict_emotion_from_audio(audio_input)
    transcribed_text = transcribe(audio_input)
    texto_imagen = emotion_prediction + transcribed_text
    image = generate_image(api_key, texto_imagen)
    return emotion_prediction, transcribed_text, image

# Create the Gradio interface
interface = gr.Interface(
    fn=get_predictions,
    inputs=gr.Audio(label="Input Audio", type="filepath", sources=["microphone"]), 
    outputs=[
        gr.Label("Acoustic Prediction", label="Acoustic Prediction"),
        gr.Label("Transcribed Text", label="Transcribed Text"),
        gr.Image(type='pil', label="Generated Image")
    ],
    title="Terra Australis Ignota: Map Generation",
    
    description="""
    Imagine Terra Australis Ignota.
    
    Speak freely. Describe what emerges:
    a place, a presence, a climate, a memory, a distortion.
    
↳ Reference: [Terra Australis Ignota](https://territoriosvirtuales.wordpress.com/1570/05/20/terra-australis-ignota/)

    To interact:
    Click Record and speak. Then stop the recording.
    You can listen, edit if needed, and submit your audio file to compose a map.
    """
    
    , 
    flagging_mode="never"  # ✅ this removes the flag button
)


interface.launch(server_name="0.0.0.0", server_port=7860)