app.py

import gradio as gr
import torch
import numpy as np
from transformers import Wav2Vec2FeatureExtractor, Wav2Vec2ForSequenceClassification
import librosa

# Initialize model and processor
device = "cuda" if torch.cuda.is_available() else "cpu"
model_name = "Hatman/audio-emotion-detection"
feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained(model_name)
model = Wav2Vec2ForSequenceClassification.from_pretrained(model_name)
model.to(device)

# Define emotion labels
EMOTION_LABELS = {
    0: "angry",
    1: "disgust",
    2: "fear",
    3: "happy",
    4: "neutral",
    5: "sad",
    6: "surprise"
}

def preprocess_audio(audio, target_sr=16000):
    """Enhanced audio preprocessing"""
    try:
        # Convert to numpy array and ensure float32
        audio = np.array(audio, dtype=np.float32)
        
        # Convert to mono if stereo
        if len(audio.shape) > 1:
            audio = librosa.to_mono(audio.T)
            
        # Resample if needed
        if target_sr != 16000:
            audio = librosa.resample(audio, orig_sr=target_sr, target_sr=16000)
            
        # Apply preprocessing steps
        # 1. Noise reduction
        audio = librosa.effects.preemphasis(audio)
        
        # 2. Normalize
        audio = librosa.util.normalize(audio)
        
        # 3. Voice activity detection
        intervals = librosa.effects.split(audio, top_db=20)
        if len(intervals) > 0:
            audio = np.concatenate([audio[start:end] for start, end in intervals])
        
        # 4. Ensure minimum length (1 second)
        if len(audio) < 16000:
            audio = np.pad(audio, (0, 16000 - len(audio)))
            
        # 5. Take center 3 seconds if too long
        if len(audio) > 48000:  # 3 seconds at 16kHz
            center = len(audio) // 2
            start = center - 24000
            end = center + 24000
            audio = audio[start:end]
            
        return audio
        
    except Exception as e:
        print(f"Preprocessing error: {str(e)}")
        return None

def get_emotion_history():
    """Get emotion detection history"""
    if not hasattr(get_emotion_history, "history"):
        get_emotion_history.history = []
    return get_emotion_history.history

def process_audio(audio):
    """Process audio chunk and return emotion"""
    if audio is None:
        return "No audio input detected"
        
    try:
        # Get the audio data
        if isinstance(audio, tuple):
            audio = audio[1]
            
        # Preprocess audio
        processed_audio = preprocess_audio(audio)
        if processed_audio is None:
            return "Audio preprocessing failed"
            
        if np.max(np.abs(processed_audio)) < 0.01:
            return "Audio too quiet"
            
        # Prepare input for the model
        inputs = feature_extractor(
            processed_audio,
            sampling_rate=16000,
            return_tensors="pt",
            padding=True
        )
        
        # Move to device and ensure float32
        inputs = {k: v.to(device, dtype=torch.float32) for k, v in inputs.items()}
        
        # Get prediction
        with torch.no_grad():
            outputs = model(**inputs)
            logits = outputs.logits
            probs = torch.nn.functional.softmax(logits, dim=-1)[0]
            
            # Get top 2 predictions
            top2_probs, top2_ids = torch.topk(probs, 2)
            
            # Convert to percentages
            top2_probs = [p * 100 for p in top2_probs.cpu().numpy()]
            top2_emotions = [EMOTION_LABELS[idx.item()] for idx in top2_ids]
            
            # Update history
            history = get_emotion_history()
            history.append(top2_emotions[0])
            if len(history) > 5:
                history.pop(0)
                
            # Get most common emotion in history
            if len(history) >= 3:
                from collections import Counter
                most_common = Counter(history).most_common(1)[0][0]
            else:
                most_common = top2_emotions[0]
                
            result = f"Primary: {top2_emotions[0]} ({top2_probs[0]:.1f}%)\n"
            result += f"Secondary: {top2_emotions[1]} ({top2_probs[1]:.1f}%)\n"
            result += f"Trending: {most_common}"
            
            return result
            
    except Exception as e:
        print(f"Error in processing: {str(e)}")
        return "Processing error. Please try again."

# Create Gradio interface
demo = gr.Interface(
    fn=process_audio,
    inputs=[
        gr.Audio(
            sources=["microphone"],
            type="numpy",
            streaming=True,
            label="Speak into your microphone",
            show_label=True
        )
    ],
    outputs=gr.Textbox(
        label="Detected Emotions",
        lines=3
    ),
    title="Enhanced Live Emotion Detection",
    description="Speak naturally into your microphone. Shows primary and secondary emotions with confidence levels.",
    live=True,
    allow_flagging=False
)

# Launch with a small queue for better real-time performance
if __name__ == "__main__":
    demo.queue(max_size=1).launch(share=True)