APEX: Large-Scale Multi-Task Aesthetic-Informed Popularity Prediction for AI-Generated Music

APEX is the first large-scale multi-task learning framework for jointly predicting popularity and aesthetic quality of AI-generated music from audio alone. Trained on over 211k AI-generated songs (~10k hours) from Suno and Udio using MERT-v1-95M audio embeddings.

---

What does APEX predict?

Given any audio file, APEX predicts 7 scores:

Score	Range	Description
score_streams	0–100	Predicted streaming engagement score
score_likes	0–100	Predicted likes engagement score
coherence	1–5	Structural and harmonic coherence
musicality	1–5	Overall musical quality
memorability	1–5	How memorable the song is
clarity	1–5	Clarity of production and mix
naturalness	1–5	Naturalness of the generated audio

---

Installation

git clone https://github.com/amaai-lab/apex.git
cd apex
pip install -r requirements.txt

---

Inference

python inference.py \
    --audio /path/to/your/mp3/file \
    --checkpoint models/best_model.pt \
    --save_json results.json

The model is also available on 🤗 HuggingFace at amaai-lab/apex.

---

Training

Step 1: Extract Embeddings

Extract MERT embeddings from your audio files:

python extract_embeddings.py \
    --jsonl_file      your_songs.jsonl \
    --audio_folder    audio_files/ \
    --parquet_folder  parquets/ \
    --num_gpus        4 \
    --songs_per_batch 50

The input JSONL file should have one song per line with ids which maps to the the mp3 file and other required label fields.

Embeddings are saved as Apache Parquet shards in the output folder.

---

Step 2: Train

The training script supports all experimental conditions via command-line arguments:

python train.py \
    --mode    [segment|song] \
    --task    [popularity|full] \
    --loss    [equal|weighted|uncertainty] \
    --shared  [2|3]

Argument	Options	Description
--mode	segment, song	Segment-level or song-level training
--task	popularity, full	2 branches (streams+likes) or 7 branches (+ aesthetics)
--loss	equal, weighted, uncertainty	Loss combination strategy
--shared	2, 3	Number of shared FC layers

This yields 24 experimental conditions (2 × 2 × 3 × 2). To train the best performing configuration:

python train.py --mode song --task full --loss uncertainty --shared 2

Checkpoints are saved to: checkpoints/loss-{loss}_shared-{shared}_mode-{mode}_task-{task}/best_model.pt

---

Step 3: Evaluate

python eval.py \
    --checkpoint    checkpoints/loss-uncertainty_shared-2_mode-song_task-full/best_model.pt \
    --test_folder   final_split/test \
    --results_folder eval_results

---

Music Arena Experiment

The notebook APEX_arena_experiment.ipynb contains the out-of-distribution pairwise human preference experiment on the Music Arena dataset, evaluating APEX predictions against human preference judgements across eleven generative music systems unseen during training.

---

License

This project is licensed under the Apache 2.0 License.