apollodB is an AI system that combines cutting-edge music emotion recognition with personalized audio engineering. Analyze your music's emotional characteristics and generate custom EQ curves tailored to your listening preferences.
Note on architecture: The primary runtime is now a FastAPI backend (backend/server.py) with a JavaScript frontend served from web/. A legacy Streamlit app (app.py) remains for reference and parity testing.
- Deep Emotional Analysis: AI-powered emotion recognition using the DEAM dataset
- Multi-Song Analysis: Upload and analyze multiple tracks for aggregate insights
- Personalized EQ Generation: Custom equalizer curves based on emotional analysis
- Multiple EQ Formats: Wavelet, Parametric, and Graphic EQ export options
- Aggression Control: Adjustable intensity for EQ modifications
- Valence-Arousal Space: Interactive 2D emotional mapping
- Emotion Distribution: Pie charts showing emotional breakdown of your music
- EQ Frequency Response: Visual representation of generated equalizer curves
- Statistical Analysis: Comprehensive metrics and insights
- Squig.link Database: Popular IEM frequency response information
- Smart Recommendations: IEM suggestions based on your music's emotional profile
- Compatibility Guide: Understanding how different IEMs complement your taste
- Spotify Integration: Direct links to curated playlists matching your emotional profile
- Mix Suggestions: Automated playlist recommendations based on dominant emotions
- Training Dataset: DEAM (Database for Emotional Analysis in Music)
- Feature Extraction: Log-mel spectrograms (128 mel bands, 1300 time frames)
- Neural Network: Convolutional layers with attention mechanisms
- Emotion Categories: Happy, Sad, Angry, Calm, Neutral (with bias correction)
- Supported Formats: MP3, WAV, M4A, FLAC
- Sample Rate: 22,050 Hz
- Duration: 30-second analysis windows
- Feature Normalization: Standardized using training data statistics
- Frequency Range: 20 Hz to 20 kHz (ISO 1/3 octave bands)
- Psychoacoustic Basis: Research-backed frequency-emotion correlations
- Customizable Intensity: User-controlled aggression parameter
- Clone the repository:
git clone <repository-url>
cd apollodB- Install dependencies:
pip install -r requirements.txt-
Ensure model files are present:
best_model.h5- Trained emotion recognition modelscaler_mean.npy- Feature normalization parametersscaler_scale.npy- Feature scaling parameterslabels.json- Emotion label mappings
-
Run the application (legacy Streamlit):
streamlit run app.py- Navigate to the "Analysis" tab
- Upload one or more audio files
- Adjust the EQ aggression slider (0.0 = subtle, 1.0 = dramatic)
- Select your preferred EQ format
- Click "Analyze Music" to process your files
- Valence: Emotional positivity (0 = negative, 1 = positive)
- Arousal: Energy level (0 = calm, 1 = energetic)
- Dominant Emotion: Most prevalent emotion across your music
- EQ Curve: Frequency adjustments optimized for your emotional profile
GraphicEQ: 20 -0.1; 21 -0.0; 22 -0.0; ...
Low Shelf (100Hz): 1.5dB
Low Mid (500Hz): 0.5dB
Mid (2kHz): 2.0dB
High Mid (8kHz): 3.0dB
High Shelf (16kHz): 2.5dB
- Trained on 4 primary emotion classes (excluding some neutral samples)
- Neutral class bias: Results with <93% confidence focus on non-neutral emotions
- Model outputs top 2 dominant emotions when neutral confidence is low
- Optimized for 30-second audio segments
- Best performance on high-quality audio files
- Robust to various musical genres and styles
- Color Palette: Black, grey, white, cyan only
- Typography: Roboto Bold throughout
- Theme: Cyberpunk meets Greek classicism
- Layout: Clean, modern, professional
- Intuitive interface with progressive disclosure
- Real-time feedback and loading indicators
- Comprehensive help and documentation
- Mobile-responsive design
apollodB is built on peer-reviewed research in:
- Music Information Retrieval (MIR)
- Affective Computing
- Psychoacoustics
- Audio Engineering
- Machine Learning
See the "References" tab in the application for detailed citations.
Q: How accurate is the emotion detection? A: The model achieves competitive performance on the DEAM benchmark. However, individual perception of musical emotion can be subjective.
Q: Why focus on valence-arousal instead of basic emotions? A: The valence-arousal model provides a more nuanced, continuous representation of emotional space, allowing for better EQ personalization.
Q: Can I use the EQ settings with any audio equipment? A: Yes, the exported EQ curves are compatible with most software and hardware equalizers that support the respective formats.
Q: Is my music data stored or shared? A: No. All processing is done locally, and temporary files are immediately deleted after analysis.
Q: Does apollodB require an internet connection? A: Only for the Spotify integration feature. All emotion analysis and EQ generation work offline.
We welcome contributions! Areas of interest:
- Additional emotion models
- New EQ curve generation algorithms
- Enhanced visualizations
- Mobile app development
- Extended IEM database
All Rights Reserved - This software is proprietary and protected by copyright law.
Copyright (c) 2025 Parikshit Kumar. All rights reserved.
Unauthorized copying, distribution, or use is strictly prohibited. For commercial licensing inquiries, contact: [email protected]
See LICENSE for complete terms.