niallone/melodygenerator

Melody Generator

AI models for generating original musical melodies, trained on MIDI data across multiple genres. These models power melodygenerator.fun.

Models

Version	Architecture	Genre	Vocab Size	Training Data	Format
v2	LSTM (3-layer)	R&B / 90s Hip Hop	59	24 songs	PyTorch + Keras
v3	LSTM (3-layer)	Dance	635	~200 songs	PyTorch + Keras
v4	LSTM (3-layer)	Jazz	635	~120 songs	PyTorch + Keras
v5	LSTM (3-layer)	Various	629	275 songs	PyTorch + Keras
v6	LSTM + Embeddings	Various	362	275 songs	PyTorch
v7	Transformer	Various	512	275 songs	PyTorch
v8	Transformer	Various + Classical	1024	275 songs + MAESTRO	PyTorch

Architecture Details

LSTM Models (v2-v6)

• 3-layer LSTM with configurable hidden sizes (256/512)
• v2-v5: Legacy pitch-string tokenization
• v6: REMI tokenization with learned embeddings (128-dim) and optional attention

Transformer Models (v7-v8)

• 8 layers, 8 attention heads, 512-dim embeddings, 2048-dim feed-forward
• RoPE (Rotary Position Embeddings)
• RMSNorm (Root Mean Square Layer Normalization)
• SwiGLU activation in feed-forward layers
• KV-cache support for efficient autoregressive generation
• REMI tokenization with BPE compression (v7: 512 vocab, v8: 1024 vocab)

File Structure

Each version directory contains:

• model.pt — PyTorch model weights
• metadata.json — Model configuration and vocabulary info
• seeds.json — Pre-computed seed sequences for generation
• tokenizer/tokenizer.json — MidiTok REMI tokenizer (v6-v8 only)
• model.h5 / model.h5_data.pkl — Legacy Keras weights (v2-v5 only)

Usage

import torch

# Load a model
checkpoint = torch.load("v8/model.pt", map_location="cpu")

For full inference with generation, MIDI conversion, and audio rendering, see the melodygenerator repository.

Links

• Live demo: melodygenerator.fun
• Source code: github.com/niallone/melodygenerator