Attention‑Enhanced Convolutional Neural Network for Music Genre Classification

M Pushparajan; KT Sreekumar; KI Ramachandran; C Santhosh Kumar

doi:10.5334/tismir.268

Attention‑Enhanced Convolutional Neural Network for Music Genre Classification

Transactions of the International Society for Music Information Retrieval

Volume 9 (2026): Issue 1

By: M Pushparajan, KT Sreekumar, KI Ramachandran and C Santhosh Kumar

Open Access

|Aug 2026

Figures & Tables

Table 1

GTZAN dataset statistics.

Segment Duration	Segments per 30‑s Track	Total Segments	Training (80%)	Testing (20%)
30 s	1	1000	800	200
15 s	2	2000	1600	400
10 s	3	3000	2400	600
6 s	5	5000	4000	1000
3 s	10	10000	8000	2000

Table 2

IMG‑Amrita dataset statistics.

Segment Duration	Segments per 30‑s Track	Total Segments	Training (80%)	Testing (20%)
30 s	1	1800	1440	360
15 s	2	3600	2880	720
10 s	3	5400	4320	1080
6 s	5	9000	7200	1800
3 s	10	18000	14400	3600

Architecture of the convolutional neural network.

Convolutional neural network—long‑term short memory network architecture.

The architecture and training steps of the proposed model.

Algorithm 1:

Proposed AECNN Framework

Require: Audio dataset $D = {(A_{i}, y_{i})}_{i = 1}^{N}$ , segment duration $T_{s}$ , sampling frequency $F_{s}$

Ensure: Trained classifier $M$

1: Initialize training set $T \leftarrow \emptyset$

2: for each audio recording $A_{i} \in D$ do

3: Segment $A_{i}$ into fixed‑duration clips

4: Apply data augmentation:

Time stretching

Pitch shifting

Noise addition

Dynamic range compression

5: for each augmented segment $a_{i j}$ do

6: Compute STFT $(a_{i j})$

7: Generate log‑mel spectrogram $S_{i j}$

8: Resize $S_{i j}$ to input dimension $H \times W \times C$

9: $T \leftarrow T \cup {(S_{i j}, y_{i})}$

10: end for

11: end for

12: Split $T$ into training set $T_{train}$ and testing set $T_{test}$

13: Initialize model parameters $θ$

14: while stopping criterion not satisfied do

15: Sample mini‑batch $B \subset T_{train}$

16: for each spectrogram $S_{k} \in B$ do

17: $F_{1} \leftarrow C N N L a y e r_{1} (S_{k})$

18: $F_{2} \leftarrow C N N L a y e r_{2} (F_{1})$

19: $P \leftarrow P a t c h E m b e d d i n g (F_{2})$

20: $E \leftarrow P + P o s i t i o n a l E n c o d i n g (P)$

21: $H_{1} \leftarrow A t t e n t i o n H e a d_{1} (E)$

22: $H_{2} \leftarrow A t t e n t i o n H e a d_{2} (E)$

23: $H \leftarrow C o n c a t e n a t e (H_{1}, H_{2})$

24: $H \leftarrow L a y e r N o r m a l i z a t i o n (H)$

25: $H \leftarrow F e e d F o r w a r d N e t w o r k (H)$

26: $H \leftarrow R e s i d u a l C o n n e c t i o n (H)$

27: $z \leftarrow F u l l y C o n n e c t e d (H)$

28: $p \leftarrow S o f t m a x (z)$

29: end for

30: Compute categorical cross‑entropy loss

31: Update parameters ( $θ$ ) using AdamW optimizer

32: end while

33: Evaluate classifier using:

Classification accuracy

Top‑five accuracy

Confusion matrix

Area under the receiver operating characteristic curve

34: return trained model $M$

Table 3

Statistical comparison of CNN and CNN‑LSTM on the GTZAN dataset (mean ± SD, 95% CI, paired t test).

Duration	CNN (Mean Acc. ± SD)	95% CI	CNN‑LSTM (Mean Acc. ± SD)	95% CI	p‑value
30 s	83.05 ± 0.45	[82.49, 83.61]	84.55 ± 0.48	[83.96, 85.14]	0.004
15 s	83.92 ± 0.44	[83.38, 84.46]	85.22 ± 0.47	[84.64, 85.80]	0.006
10 s	85.25 ± 0.42	[84.73, 85.77]	87.32 ± 0.50	[86.69, 87.95]	<0.001
6 s	86.13 ± 0.41	[85.62, 86.64]	88.45 ± 0.52	[87.80, 89.10]	<0.001
3 s	87.05 ± 0.40	[86.56, 87.54]	89.63 ± 0.55	[88.94, 90.32]	<0.001

Table 4

Statistical comparison of CNN and CNN‑LSTM on the IMG‑Amrita dataset (mean ± SD, 95% CI, paired t test).

Duration	CNN (Mean Acc. ± SD)	95% CI	CNN‑LSTM (Mean Acc. ± SD)	95% CI	p‑value
30 s	90.80 ± 0.43	[90.27, 91.33]	91.52 ± 0.46	[90.95, 92.09]	0.041
15 s	91.45 ± 0.41	[90.94, 91.96]	91.89 ± 0.44	[91.35, 92.43]	0.083
10 s	92.05 ± 0.39	[91.56, 92.54]	93.11 ± 0.47	[92.53, 93.69]	0.009
6 s	93.23 ± 0.38	[92.76, 93.70]	95.05 ± 0.49	[94.44, 95.66]	<0.001
3 s	95.55 ± 0.36	[95.10, 96.00]	96.67 ± 0.42	[96.15, 97.19]	0.003

Table 5

Training hyperparameters of the CNN and CNN‑LSTM models.

Hyperparameter	Value
Optimizer	Adam
Learning rate	0.001
Loss function	Categorical cross‑entropy
Batch size	128
Number of epochs	100
Early stopping patience	10
Learning rate reduction patience	5
Learning rate reduction factor	0.5
Minimum learning rate	1e‑5

Table 6

Statistical erformance valuation of the AECNN‑based MGC system (GTZAN) (mean ± SD, 95% CI, paired t test).

Duration	AECNN (Mean Acc. SD)	95% CI	p‑value vs. CNN	p‑value vs. CNN‑LSTM
30 s	90.68 ± 0.43	[90.14, 91.22]	0.002	0.006
15 s	91.12 ± 0.45	[90.56, 91.68]	0.001	0.004
10 s	92.05 ± 0.46	[91.48, 92.62]	<0.001	0.002
6 s	93.82 ± 0.48	[93.22, 94.42]	<0.001	<0.001
3 s	96.02 ± 0.51	[95.39, 96.65]	<0.001	<0.001

Table 7

Statistical performance evaluation of the AECNN‑based MGC system (IMG‑Amrita) (mean ± SD, 95% CI, paired t test).

Duration	AECNN (Mean Acc. SD)	95% CI	p‑value vs. CNN	p‑value vs. CNN‑LSTM
30 s	93.62 ± 0.40	[93.12, 94.12]	0.003	0.021
15 s	94.68 ± 0.42	[94.15, 95.21]	0.002	0.014
10 s	95.94 ± 0.44	[95.39, 96.49]	0.001	0.006
6 s	98.21 ± 0.47	[97.62, 98.80]	<0.001	0.002
3 s	98.94 ± 0.46	[98.36, 99.52]	<0.001	<0.001

Confusion matrix of the attention‑enhanced convolutional neural network–based music genre classification system (3 s) (Western genre).

Area under the receiver operating characteristic curve analysis of the attention‑enhanced convolutional neural network–based music genre classification system (3 s) (Western genre).

Confusion matrix of the attention‑enhanced convolutional neural network–based music genre classification system (3 s) (Indian genre).

Table 8

Performance comparison of the CNN and CNN‑LSTM with the proposed method.

Genre	Method	Accuracy (%)	Precision (%)	Recall (%)	F1‑Score (%)
Western	CNN	87.05	87.65	87.20	86.40
	CNN‑LSTM	89.63	90.03	89.55	88.74
	Proposed method	96.02	96.48	95.63	96.05
Indian	CNN	95.55	95.90	94.87	94.34
	CNN‑LSTM	96.67	97.12	96.05	95.40
	Proposed method	98.94	99.12	98.63	98.87

Table 9

Training hyperparameters of the proposed AECNN model.

Hyperparameter	Value
Optimizer	AdamW
Learning rate	$1 \times 1 0^{- 3}$
Weight decay	$1 \times 1 0^{- 4}$
Batch size	128
Number of epochs	100
Loss function	Categorical cross‑entropy
Label smoothing	0.1
Dropout rate	0.1
Stochastic depth rate	0.01

Table 10

Results of the ablation study (Western genre).

Configuration	Description	Accuracy (%) Mean ± SD [95% CI]	Precision (%) Mean ± SD [95% CI]	Recall (%) Mean ± SD [95% CI]	F1 Score Mean ± SD [95% CI]
Without batch normalization	Removed batch normalization from convolutional layers	93.53 ± 0.38 [93.20, 93.86]	94.08 ± 0.35 [93.77, 94.39]	92.87 ± 0.41 [92.51, 93.23]	93.47 ± 0.37 [93.15, 93.79]
Reduced CNN depth	Reduced the depth of the CNN backbone (from 3 to 2)	92.79 ± 0.42 [92.42, 93.16]	93.41 ± 0.39 [93.07, 93.75]	92.06 ± 0.44 [91.67, 92.45]	92.73 ± 0.40 [92.38, 93.08]
Reduced attention heads	Reduced the number of attention heads (from 2 to 1)	92.03 ± 0.46 [91.63, 92.43]	92.68 ± 0.43 [92.30, 93.06]	91.42 ± 0.48 [91.00, 91.84]	92.04 ± 0.45 [91.65, 92.43]
Increased attention heads	Increased the number of attention heads (from 2 to 3)	94.98 ± 0.34 [94.68, 95.28]	95.37 ± 0.31 [95.10, 95.64]	94.52 ± 0.36 [94.20, 94.84]	94.94 ± 0.33 [94.65, 95.23]
Without attention module	Removed the attention module; only CNN remains	89.11 ± 0.51 [88.66, 89.56]	90.02 ± 0.48 [89.60, 90.44]	88.27 ± 0.55 [87.79, 88.75]	89.13 ± 0.50 [88.69, 89.57]
Proposed AECNN	Complete architecture (CNN + dual attention modules)	96.02 ± 0.29 [95.76, 96.28]	96.48 ± 0.27 [96.24, 96.72]	95.63 ± 0.31 [95.36, 95.90]	96.05 ± 0.28 [95.80, 96.30]

Table 11

Results of the ablation study (Indian genre).

Configuration	Description	Accuracy (%) Mean ± SD [95% CI]	Precision (%) Mean ± SD [95% CI]	Recall (%) Mean ± SD [95% CI]	F1 Score Mean ± SD [95% CI]
Without batch normalization	Removed batch normalization from convolutional layers	97.53 ± 0.27 [97.29, 97.77]	97.88 ± 0.25 [97.66, 98.10]	97.21 ± 0.30 [96.95, 97.47]	97.54 ± 0.28 [97.29, 97.79]
Reduced CNN depth	Reduced the depth of the CNN backbone (from 3 to 2)	97.01 ± 0.31 [96.74, 97.28]	97.36 ± 0.29 [97.11, 97.61]	96.68 ± 0.33 [96.39, 96.97]	97.02 ± 0.30 [96.76, 97.28]
Reduced attention heads	Reduced the number of attention heads (from 2 to 1)	96.28 ± 0.36 [95.96, 96.60]	96.71 ± 0.34 [96.41, 97.01]	95.83 ± 0.39 [95.49, 96.17]	96.27 ± 0.35 [95.96, 96.58]
Increased attention heads	Increased the number of attention heads (from 2 to 3)	97.88 ± 0.24 [97.67, 98.09]	98.21 ± 0.22 [98.02, 98.40]	97.46 ± 0.27 [97.22, 97.70]	97.83 ± 0.25 [97.61, 98.05]
Without attention module	Removed the attention module; only CNN remains	96.00 ± 0.40 [95.65, 96.35]	96.38 ± 0.37 [96.06, 96.70]	95.62 ± 0.42 [95.25, 95.99]	95.99 ± 0.39 [95.65, 96.33]
Proposed AECNN	Complete architecture (CNN + dual attention modules)	98.94 ± 0.18 [98.78, 99.10]	99.12 ± 0.16 [98.98, 99.26]	98.63 ± 0.21 [98.45, 98.81]	98.87 ± 0.19 [98.70, 99.04]

Table 12

Comparison of the Proposed AECNN Model with the State‑of‑the‑Art Models for the GTZAN Dataset.

Approach	Accuracy (%)	Precision (%)	Recall (%)	F1‑Score (%)
Hybrid CNN‑RNN (Ashraf et al., 2023)	89.30	85.00	91.00	88.00
CNN–transformer encoder (Chen et al., 2024)	87.41	87.93	87.58	87.28
Compact convolutional transformer (Pourmoazemi and Maleki, 2024)	94.00	92.00	93.00	92.00
Proposed work (AECNN)	96.02	96.48	95.63	96.05

Radar chart comparison of the proposed attention‑enhanced convolutional neural network model with the state‑of‑the‑art models for GTZAN dataset.

References

Authors

Metrics

Articles in this issue

DOI: https://doi.org/10.5334/tismir.268 | Journal eISSN: 2514-3298

Journal RSS Feed

Language: English

Page range: 423 - 439

Submitted on: Apr 10, 2025

Accepted on: Jun 18, 2026

Published on: Aug 4, 2026

Published by: Ubiquity Press

In partnership with: Paradigm Publishing Services

Keywords:

Music genre classification,

Music information retrieval,

Convolutional neural networks,

Mel‑spectrogram,

Attention mechanism

© 2026 M Pushparajan, KT Sreekumar, KI Ramachandran, C Santhosh Kumar, published by Ubiquity Press
This work is licensed under the Creative Commons Attribution 4.0 License.

Volume 9 (2026): Issue 1

Attention‑Enhanced Convolutional Neural Network for Music Genre Classification

Figures & Tables

Table 1

Table 2

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Algorithm 1:

Table 3

Table 4

Table 5

Table 6

Table 7

Figure 6

Figure 7

Figure 8

Figure 9

Table 8

Table 9

Table 10

Table 11

Table 12

Figure 10

Paradigm

My account