wav2vec2-base-is_vinyl_scratched_or_not

This model is a fine-tuned version of facebook/wav2vec2-base on the audiofolder dataset. It achieves the following results on the evaluation set:

Loss: 0.1039
Accuracy: 0.9752
F1: 0.9638
Recall: 0.9576
Precision: 0.9700

Model description

This is a binary classifier that predicts whether or not the vinyl record played in the audio sample is scratched.

For more information on how it was created, check out the following link: https://github.com/DunnBC22/Vision_Audio_and_Multimodal_Projects/blob/main/Audio-Projects/Classification/Vinyl%20Scratched%20or%20Not.ipynb

Intended uses & limitations

This model is intended to demonstrate my ability to solve a complex problem using technology.

Training and evaluation data

Dataset Source: https://www.kaggle.com/datasets/seandaly/detecting-scratch-noise-in-vinyl-playback

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 3e-05
train_batch_size: 32
eval_batch_size: 32
seed: 42
gradient_accumulation_steps: 4
total_train_batch_size: 128
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
lr_scheduler_warmup_ratio: 0.1
num_epochs: 10

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy	F1	Recall	Precision
0.6671	0.98	21	0.6235	0.6560	0.0	0.0	0.0
0.4954	1.98	42	0.2824	0.9417	0.9095	0.8517	0.9757
0.2406	2.98	63	0.1755	0.9563	0.9336	0.8941	0.9769
0.169	3.98	84	0.1545	0.9592	0.9386	0.9068	0.9727
0.1287	4.98	105	0.1249	0.9606	0.9407	0.9068	0.9772
0.1102	5.98	126	0.1159	0.9723	0.9595	0.9534	0.9657
0.0923	6.98	147	0.1073	0.9665	0.9516	0.9576	0.9456
0.0877	7.98	168	0.1039	0.9752	0.9638	0.9576	0.9700
0.0807	8.98	189	0.1088	0.9679	0.9536	0.9576	0.9496
0.0744	9.98	210	0.1041	0.9752	0.9638	0.9576	0.9700

Framework versions

Transformers 4.26.0
Pytorch 1.12.1
Datasets 2.8.0
Tokenizers 0.12.1