Filharmonia AST - Concert Audio Classifier

Pre-trained Audio Spectrogram Transformer for classifying concert recordings.

Model: Audio Spectrogram Transformer (MIT/PSLA) Trained on: Classical philharmonic concert recordings Test accuracy: 97.75%

Classes

  • MUSIC - orchestral music performance
  • APPLAUSE - audience applause (marks piece boundaries)
  • SPEECH - announcements, introductions
  • PUBLIC - crowd noise, intermissions
  • TUNING - orchestra tuning

Usage

Option 1: With Hugging Face Transformers

import torch
from transformers import ASTFeatureExtractor, ASTForAudioClassification

# Download model
model = ASTForAudioClassification.from_pretrained("szymontex/filharmonia-ast")
feature_extractor = ASTFeatureExtractor.from_pretrained("MIT/ast-finetuned-audioset-10-10-0.4593")

# Your audio processing code here

Option 2: With Filharmonia AI

Use with Filharmonia AI:

  1. Download ast_20251009_222204.pth
  2. Place in RECOGNITION_MODELS/ast_active.pth
  3. Run analysis through web UI

Model Details

Architecture:

  • Base: MIT/PSLA Audio Spectrogram Transformer
  • Fine-tuned on concert recordings
  • Input: 16kHz audio, 2.97s segments
  • Features: Fbank (filter bank), 1024 time frames × 128 freq bins

Training:

  • Dataset: ~1200 min of classical concert recordings
  • Split: 80% train / 10% val / 10% test
  • Batch size: 32
  • Epochs: 10
  • Balance strength: 0.75

Performance:

Class Accuracy
APPLAUSE 100%
MUSIC 100%
SPEECH 100%
PUBLIC 96.2%
TUNING 85.7%
Overall 97.75%

Limitations

  • Trained on classical orchestral concerts only
  • May not work well on rock, jazz, pop, or other genres
  • Requires retraining for different audio types
  • Best results with similar recording conditions (concert hall acoustics)

Training Your Own

For non-classical music or different use cases:

  1. Clone filharmonia-ai
  2. Collect your audio samples
  3. Export segments to training data
  4. Retrain via web UI

See Development Guide for details.

License

MIT - same as code repository

Citation

@software{filharmonia_ast,
  author = {Szymon Gwóźdź},
  title = {Filharmonia AST: Concert Audio Classifier},
  year = {2025},
  url = {https://github.com/szymontex/filharmonia-ai}
}

Links

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support