ast-urbansound8k-finetuned-fold10

AST fine-tuned on UrbanSound8K (fold10 protocol)

This model is an ASTForAudioClassification checkpoint fine-tuned from:

• MIT/ast-finetuned-audioset-10-10-0.4593

Training protocol

• train folds: 1-8
• validation fold: 9
• test fold: 10

Validation result

• best epoch: 5
• best validation accuracy on fold 9: 0.8970588235294118

Test result (fold 10)

• samples: 837
• loss: 0.503164
• accuracy: 0.890084

Per-class accuracy (fold 10)

• air_conditioner: 0.8000 (80/100)
• car_horn: 0.9697 (32/33)
• children_playing: 0.9800 (98/100)
• dog_bark: 0.8800 (88/100)
• drilling: 0.8100 (81/100)
• engine_idling: 0.8710 (81/93)
• gun_shot: 0.9688 (31/32)
• jackhammer: 0.9896 (95/96)
• siren: 0.7711 (64/83)
• street_music: 0.9500 (95/100)

Labels

air_conditioner, car_horn, children_playing, dog_bark, drilling, engine_idling, gun_shot, jackhammer, siren, street_music

Usage

import torch
from transformers import AutoFeatureExtractor, ASTForAudioClassification

model_id = "Adam-ousse/ast-urbansound8k-finetuned-fold10"
feature_extractor = AutoFeatureExtractor.from_pretrained(model_id)
model = ASTForAudioClassification.from_pretrained(model_id)
model.eval()

# waveform: 1D float32 numpy array at 16 kHz
# inputs = feature_extractor([waveform], sampling_rate=16000, return_tensors="pt", padding=True)
# with torch.no_grad():
#     logits = model(**inputs).logits
#     pred = int(torch.argmax(logits, dim=1)[0])

Notes

• This repository keeps the fixed UrbanSound8K protocol: train 1-8, val 9, test 10.
• Dataset licensing and terms are governed by UrbanSound8K.