GenreBlender: Generative AI Music Mixer & Classifier

Inspiration

Music genres aren’t fixed — they constantly blend and evolve. We wanted to explore what happens if you could control that blending in real time. Instead of relying only on prompt engineering, we asked: how can we measure whether a genre blend actually worked?

---

What it does

GenreBlender lets users:

• Select two genres
  
• Adjust a blending slider (α ∈ [0,1])
  
• Generate a new AI-composed track
  
• See how closely the result matches the intended blend
  

We define the target blend as:

Target = (α · A) + ((1 − α) · B)

Our classifier predicts:

P(genre_i | audio)

We compare the predicted distribution to the target, making genre blending measurable and controllable.

---

How we built it

1. Generative Engine
We used Meta’s MusicGen to generate tracks from weighted prompts like:
“70% classical, 30% hip hop.”

2. Neural Genre Classifier

• Trained for 100 epochs on the GTZAN dataset using a Multilayer Perceptron with 4 layers (3 hidden, 1 output), achieving a 92% accuracy onthe Validation Set
• Extracted MFCCs, RMS, harmonic/percussive features, tempo
  
• Used StandardScaler + LabelEncoder
  
• Prevented leakage with GroupShuffleSplit
  

MLP architecture:

nn.Linear(input_size, 256)
ReLU
Dropout(0.3)

nn.Linear(256, 128)
ReLU
Dropout(0.3)

nn.Linear(128, 64)
ReLU
Dropout(0.3)

nn.Linear(64, num_classes)

Pipeline:

1. User selects genres + α
   
2. MusicGen generates audio
   
3. Features extracted
   
4. Classifier predicts genre probabilities
   
5. Results returned to frontend
   

---

Challenges we ran into

• CUDA/GPU compatibility issues
  
• Ensuring identical preprocessing during training and inference
  
• Preventing data leakage in audio splits
  
• Optimizing generation speed for real-time demos
  

---

Accomplishments that we're proud of

• Built both a generative model interface and a custom evaluation model
  
• Achieved 92% validation accuracy
  
• Created a closed-loop controllable generation system
  
• Made music blending interactive, interpretable, and quantifiable

---

What we learned

• Prompt engineering alone isn’t enough for controllable AI
  
• Evaluation models improve interpretability
  
• Preventing data leakage is critical
  
• Thinking in systems (generation + evaluation) leads to stronger AI applications
  

---

What's next for GenreBlender: Generative AI Music Mixer & Classifier

• Use embedding-level interpolation instead of prompt-only blending
  
• Add more genres and larger datasets
  
• Introduce divergence metrics to quantify blend accuracy
  
• Deploy a scalable GPU-hosted version
  

Built With

• joblib
• mlp
• numpy
• pandas
• python
• pytorch
• scikit-learn
• streamlit