🎤 GaMaDHaNi: Hierarchical Generative Modeling of Melodic Vocal Contours in Hindustani Classical Music

GaMaDHaNi is a modular two-level hierarchy, consisting of a generative model on pitch contours, and a pitch contour to audio synthesis model.

📖 Read our ISMIR 2024 Paper here
🎧 Check out the audio samples here
💻 Play with the interactive demo here

Installation

git clone https://github.com/snnithya/GaMaDHaNi.git
pip install -r requirements.txt

How to use

Generating without any melodic prompt (no pitch prime)

Diffusion-based Pitch Generation Model:

cd GaMaDHaNi
python generate.py --pitch_model_type=diffusion --prime=False --number_of_samples=1 --download_model_from_hf=True

Transformer-based Pitch Generation Model:

cd GaMaDHaNi
python generate.py --pitch_model_type=transformer --prime=False --number_of_samples=1 --download_model_from_hf=True

Generating with predefined melodic prompts (pitch primes)

Note: You will need download_model_from_hf=True to be able to access the pitch primes. You will be able to see the primes (first 4s of all generations) plotted in a different colour in the pitch plots of generated samples. 'num_samples' can go from 1 to 16 for generation with primes.

Diffusion-based Pitch Generation Model:

cd GaMaDHaNi
python generate.py --pitch_model_type=diffusion --prime=True --number_of_samples=1 --download_model_from_hf=True 

Transformer-based Pitch Generation Model:

cd GaMaDHaNi
python generate.py --pitch_model_type=transformer --prime=True --number_of_samples=1 --download_model_from_hf=True 

Training the Pitch Generation Model

Transformer-based Pitch Generation Model:

cd GaMaDHaNi
python gamadhani/scripts/train_transformer.py --config configs/transformer_pitch_config.gin --db_path HF_DB_PATH --gpu=0 --val_every=1 --max_epochs=500 --batch_size=4

Note: HF_DB_PATH is soon to be released.

Training scripts for the diffusion-based Pitch Generation Model and Pitch to Audio Generation model are soon to be released. Stay tuned!

BibTex

@article{shikarpur2024hierarchical,
  title={Hierarchical Generative Modeling of Melodic Vocal Contours in Hindustani Classical Music},
  author={Shikarpur, Nithya and Dendukuri, Krishna Maneesha and Wu, Yusong and Caillon, Antoine and Huang, Cheng-Zhi Anna},
  journal={arXiv preprint arXiv:2408.12658},
  year={2024}
}