ntua-slp/CultureMERT-95M

CultureMERT: Continual Pre-Training for Cross-Cultural Music Representation Learning

📑 Read the full paper (to be presented at ISMIR 2025)

CultureMERT-95M is a multi-culturally adapted 95M-parameter music foundation model based on MERT-v1-95M. It is developed through a two-stage continual pre-training (CPT) strategy on 650 hours of culturally diverse audio spanning Greek, Turkish, and Indian musical traditions. The model significantly improves representation quality for "non-Western" music, achieving an average improvement of +4.9% across ROC-AUC and mAP on culturally diverse non-Western music tagging tasks, surpassing prior state-of-the-art, while maintaining strong performance on Western-centric benchmarks such as MagnaTagATune and FMA-medium.

🔀 Alternative variant available: CultureMERT-TA-95M is a merged model constructed via task arithmetic, combining separately adapted models—each continually pre-trained (using the same two-stage CPT strategy) on data from a single musical tradition.

🧠 Model Details

• Architecture: 12-layer Transformer encoder (768-dim) with a 7-layer 1D CNN frontend
• Input: Raw mono audio at 24kHz
• Training Context Length: 5 seconds
• Pretraining Objective: MLM-style multi-task masked prediction of discrete EnCodec acoustic tokens and continuous constant-Q transform (CQT) spectrogram reconstruction at a 75 Hz feature rate

---

🌍 Training Data

Dataset	Music Tradition	Hours Used
Lyra	Greek traditional/folk	50h
Turkish-makam	Turkish/Ottoman classical	200h
Hindustani	North Indian classical	200h
Carnatic	South Indian classical	200h

• To further stabilize adaptation and mitigate catastrophic forgetting, we include 20% Western data (i.e., 20 hours from Music4All) during Stage 1 of continual pre-training.

🔒 The datasets used were obtained under research-use agreements and are not redistributed.

---

📊 Evaluation

We evaluate CultureMERT-95M via probing on both Western and non-Western music auto-tagging tasks. All results are averaged over five random seeds. Metrics used:

• ROC-AUC (Area Under the Receiver Operating Characteristic Curve)
• mAP (Mean Average Precision)
• Micro-F1 and Macro-F1

Evaluation follows the MARBLE protocol under constrained settings. We use standardized train/test splits from ccml for continual pre-training and probing-based evaluation.

Evaluation Datasets and Metadata Used (Top-k Tags)

• Non-Western traditions:
  • Turkish-makam: Top-30 tags, covering makam, usul, and instruments
  • Hindustani: Top-20 tags, primarily reflecting raga, tala, instruments, and forms
  • Carnatic: Top-20 tags, primarily reflecting raga, tala, instruments, and forms
  • Lyra: Top-30 tags, related to genre, regional (place), and instruments metadata
• Western benchmarks:
  • MagnaTagATune (MTAT): Top-50 tags, spanning genre, instruments, and mood
  • FMA-medium: Top-20 hierarchical genre tags

Top-k tags were selected based on tag frequency distributions.

The evaluation results are shown in the following tables:

ROC-AUC / mAP

	Turkish-makam	Hindustani	Carnatic	Lyra	FMA-medium	MTAT	Avg.
MERT-v1-95M	83.2% / 53.3%	82.4% / 52.9%	74.9% / 39.7%	85.7% / 56.5%	90.7% / 48.1%	89.6% / 35.9%	66.1%
CultureMERT-95M	89.6% / 60.6%	88.2% / 63.5%	79.2% / 43.1%	86.9% / 56.7%	90.7% / 48.1%	89.4% / 35.9%	69.3%

Micro-F1 / Macro-F1

	Turkish-makam	Hindustani	Carnatic	Lyra	FMA-medium	MTAT	Avg.
MERT-v1-95M	73.0% / 38.9%	71.1% / 33.2%	80.1% / 30.0%	72.4% / 42.6%	57.0% / 36.9%	35.7% / 21.2%	49.3%
CultureMERT-95M	77.4% / 45.8%	77.8% / 50.4%	82.7% / 32.5%	73.1% / 43.1%	58.3% / 36.6%	35.6% / 22.9%	52.9%

📈 CultureMERT-95M outperforms the original MERT-v1-95M by an average of +4.4% in ROC-AUC across non-Western traditions, with consistent improvements of +5.4% in mAP, +3.6% in Micro-F1, and +6.8% in Macro-F1. On Western datasets, it maintains performance with only a −0.05% average drop across ROC-AUC and mAP, and even improves by +0.65% in Micro/Macro-F1, resulting in an overall average gain of +0.3% across all metrics and datasets (MTAT and FMA-medium).

---

🔧 Model Usage

from transformers import Wav2Vec2FeatureExtractor, AutoModel
import torch
from torch import nn
import torchaudio.transforms as T
from datasets import load_dataset

# Load model weights and preprocessor config
model = AutoModel.from_pretrained("ntua-slp/CultureMERT-95M", trust_remote_code=True)
processor = Wav2Vec2FeatureExtractor.from_pretrained("ntua-slp/CultureMERT-95M", trust_remote_code=True)

# Load example audio
dataset = load_dataset("hf-internal-testing/librispeech_asr_demo", "clean", split="validation", trust_remote_code=True).sort("id")
audio_array = dataset[0]["audio"]["array"]
sampling_rate = dataset.features["audio"].sampling_rate

# Resample if needed
resample_rate = processor.sampling_rate
if resample_rate != sampling_rate:
    print(f'Setting sample rate from {sampling_rate} to {resample_rate}')
    resampler = T.Resample(sampling_rate, resample_rate)
else:
    resampler = None

# Audio file is decoded on the fly
if resampler is None:
    input_audio = dataset[0]["audio"]["array"]
else:
  input_audio = resampler(torch.from_numpy(dataset[0]["audio"]["array"]).to(dtype=resampler.kernel.dtype))

# Extract hidden states
inputs = processor(input_audio, sampling_rate=resample_rate, return_tensors="pt")
with torch.no_grad():
    outputs = model(**inputs, output_hidden_states=True)

# Representations: 13 layers (CNN feature extractor + 12 Transformer)
# NOTE: each layer performs differently in different downstream tasks - you should choose empirically
all_layer_hidden_states = torch.stack(outputs.hidden_states).squeeze()
print(all_layer_hidden_states.shape) # [13 layers, Time steps, 768 feature_dim]

# For utterance-level classification tasks, you can simply reduce the representation in time
time_reduced_hidden_states = all_layer_hidden_states.mean(-2)
print(time_reduced_hidden_states.shape) # [13, 768]

# You can even use a learnable weighted average representation over all layers
aggregator = nn.Conv1d(in_channels=13, out_channels=1, kernel_size=1)
weighted_avg_hidden_states = aggregator(time_reduced_hidden_states.unsqueeze(0)).squeeze()
print(weighted_avg_hidden_states.shape) # [768]

---

Ethical Considerations

This model is released under a non-commercial CC BY-NC 4.0 license and is intended for research purposes. While it is designed to address cultural bias in MIR, its training data and pretraining paradigm may still reflect cultural and dataset-specific biases. The model should not be used in commercial or generative applications without explicit consideration of cultural representation, proper attribution, and consent from relevant communities or dataset curators.

📚 Citation

@misc{kanatas2025culturemertcontinualpretrainingcrosscultural,
      title={CultureMERT: Continual Pre-Training for Cross-Cultural Music Representation Learning}, 
      author={Angelos-Nikolaos Kanatas and Charilaos Papaioannou and Alexandros Potamianos},
      year={2025},
      eprint={2506.17818},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={https://arxiv.org/abs/2506.17818}, 
}

---