MNIST audio classification using Pytorch

Dataset

The original dataset is available in https://www.kaggle.com/datasets/sripaadsrinivasan/audio-mnist

CNN model

Model Summary

The following table shows the model architecture.

Model summary

Layer (type)	Output Shape	Param #
Conv2d-1	[-1, 10, 1023, 30]	260
ReLU-2	[-1, 10, 1023, 30]	0
MaxPool2d-3	[-1, 10, 511, 15]	0
Conv2d-4	[-1, 20, 509, 13]	5,020
ReLU-5	[-1, 20, 509, 13]	0
MaxPool2d-6	[-1, 20, 254, 6]	0
Flatten-7	[-1, 30480]	0
Linear-8	[-1, 3800]	115,827,800
ReLU-9	[-1, 3800]	0
Dropout-10	[-1, 3800]	0
Linear-11	[-1, 1024]	3,892,224
ReLU-12	[-1, 1024]	0
Dropout-13	[-1, 1024]	0
Linear-14	[-1, 256]	262,400
ReLU-15	[-1, 256]	0
Dropout-16	[-1, 256]	0
Linear-17	[-1, 10]	2,570
Total params: 119,990,274 Trainable params: 119,990,274 Non-trainable params: 0 Input size (MB): 0.13 Forward/backward pass size (MB): 7.87 Params size (MB): 457.73 Estimated Total Size (MB): 465.72

Results

Training parameters

Parameter	Value
Loss function	CrossEntroppyLoss
Optimizer	Adam
Learning rate	10^-3
Audio sample rate	16 KHz
Dataset size	30000 samples
Dataset split	70% train / 20% validation / 10% test
Batch size	32
Epochs	15
Early stopping	Patience = 3 epochs | Delta = 0.00 | Monitor = Accuracy
Device	CUDA GPU (NVIDIA GeForce RTX 3050 Laptop GPU)

Evaluation Results

Metric	Training	Validation	Test
Accuracy (%)	1.000	0.986 (best 0.991 at epoch 4)	0.9896
Loss	0.000139	0.0613	0.0458

Note: Effective training duration was 7 epochs — best validation at epoch 4, with early stopping patience set to 3 and the accuracy was 0.991.

Training Statistics

Training loss

Training accuracy

Validation loss

Validation accuracy

CNN + LSTM model

Model Summary

The following table shows the model architecture.

Model summary

Layer (type:depth-idx)	Output Shape	Param #
Conv2d: 2-1	[1, 20, 1025, 32]	200
ReLU: 2-2	[1, 20, 1025, 32]	--
MaxPool2d: 2-3	[1, 20, 512, 16]	--
LSTM: 1-2	[1, 4, 128]	21,038,080
Linear: 1-3	[1, 10]	1,290
Total params: 21,039,570 Trainable params: 21,039,570 Non-trainable params: 0 Total mult-adds (Units.MEGABYTES): 90.71
Input size (MB): 0.13 Forward/backward pass size (MB): 5.25 Params size (MB): 84.16 Estimated Total Size (MB): 89.54

Results

Training parameters

Parameter	Value
Loss function	CrossEntroppyLoss
Optimizer	Adam
Learning rate	10^-3
Audio sample rate	16 KHz
Dataset size	30000 samples
Dataset split	70% train / 20% validation / 10% test
Batch size	32
Epochs	15
Early stopping	Patience = 3 epochs | Delta = 0.00 | Monitor = Accuracy
Device	CUDA GPU (NVIDIA GeForce RTX 3050 Laptop GPU)

Evaluation Results

Metric	Training	Validation	Test
Accuracy (%)	1.000	0.985 (best 0.991 at epoch 5)	0.9900
Loss	0.00266	0.0434	0.0347

Note: Effective training duration was 8 epochs — best validation at epoch 5, with early stopping patience set to 3 and the accuracy was 0.991.

Training Statistics

Training loss

Training accuracy

Validation loss

Validation accuracy