Song Identification Using Audio Fingerprinting and Deep Learning

---

Table of Contents

• Part I: Theory & Methods
  • 1. Audio Fingerprinting — The Classical Approach
    • 1.1 Introduction
    • 1.2 Spectrogram Analysis & Peak Picking
    • 1.3 Pairwise Landmark Hashing
    • 1.4 Inverted-Index Retrieval
    • 1.5 Hyperparameter Optimization
  • 2. Deep Learning for Song ID — A CNN Approach
    • 2.1 Retrieval via Embeddings
    • 2.2 Contrastive Learning (InfoNCE)
    • 2.3 Input Representation: Log-Mel Spectrograms
    • 2.4 Architectures
    • 2.5 Training Strategy & Tuning
    • 2.6 Advanced Training Techniques
  • 3. State-of-the-Art with Pretrained Transformers
    • 3.1 Large-Scale Pretrained Models
    • 3.2 Audio Spectrogram Transformer (AST)
    • 3.3 Music Understanding Transformer (MERT)
• Part II: Experiments & Results
  • 1. Dataset Construction
  • 2. Classical Landmark Fingerprinting — Results
  • 3. CNN Results
  • 4. Transformer Results
  • 5. Comparisons
• Conclusions

---

Part I: Theory & Methods

1. Audio Fingerprinting — The Classical Approach

1.1 Introduction

Modern audio recognition systems rely on audio fingerprints—compact signatures that uniquely characterize a recording. Landmark-based algorithms (pioneered by Shazam) extract sparse, high-contrast features from the time–frequency plane. We implemented a Shazam-style fingerprinting system and explain below why each step yields robustness and speed.

1.2 Spectrogram Analysis & Peak Picking

We convert a waveform to a spectrogram with STFT:

• Load: y, sr = librosa.load(path, sr=None)
• STFT: D = librosa.stft(y, n_fft=2048, hop_length=512)
• Magnitude → dB: S = librosa.amplitude_to_db(np.abs(D))

Instead of using all time–frequency bins, we select only strong local maxima (“peaks”). A point (t₀, f₀) is a peak if |X(t₀, f₀)| exceeds all neighbors in a local time–frequency neighborhood.

Neighborhood sizes (time τ and frequency κ) trade off density vs. salience: larger windows → fewer but stronger peaks; smaller windows → denser peaks (bigger index, higher recall). The selected peaks form a sparse constellation map.
We also apply an energy threshold: compute per-frequency median energy and keep peaks only if their magnitude exceeds ETH * med[f], suppressing noisy artifacts.

1.3 Pairwise Landmark Hashing

A single peak (t, f) is not distinctive enough. We pair peaks: for each anchor peak, select target peaks occurring shortly after it, and hash the tuple (f_1, f_2, Δt) where

$$ \Delta t = t_2 - t_1. $$

These landmark hashes are compact and highly discriminative.

1.4 Inverted-Index Retrieval

All hashes are stored in an inverted index mapping hash → [(songID, timeOffset), ...]. At query time, we hash the incoming audio, look up matches, and compute offsets

$$ \delta t = t_{\text{db}} - t_{\text{query}}, $$

then vote for each ((\text{songID}, \delta t)). The correct song surfaces as a sharp vote cluster at a consistent offset—robust to noise and partial matches.

1.5 Hyperparameter Optimization

Key parameters and why they matter:

• Peak Neighborhood Size (PNS): window for local maxima. Smaller PNS → denser peaks (higher recall; bigger index). Larger PNS → sparser constellation (more salience).
• Fan (pairings per anchor): controls target zone density. Larger Fan (e.g., 50) yields more hashes per anchor (better recall; more collisions risk).
• Energy Threshold (ETH): filters low-energy peaks. Lower ETH (e.g., 0.3) retains only peaks above a stronger relative baseline, reducing spurious landmarks.

---

2. Deep Learning for Song ID — A CNN Approach

2.1 Retrieval via Embeddings

We formulate song ID as embedding retrieval. A model maps clips to vectors so that two segments from the same song are close (high cosine similarity), and segments from different songs are far apart. Inference computes the query embedding and performs nearest-neighbor search among stored song embeddings.

2.2 Contrastive Learning (InfoNCE)

We train with InfoNCE. For each anchor (x), a positive (x^+) is another segment from the same song; negatives (x^-_i) are from other songs in the batch. With cosine similarity sim(·,·) and temperature τ:

$$ L_{\mathrm{InfoNCE}} = -\log \frac{\exp(\mathrm{sim}(x,x^+)/\tau)} {\exp(\mathrm{sim}(x,x^+)/\tau) + \sum_i \exp(\mathrm{sim}(x,x^-_i)/\tau)}. $$

Smaller τ sharpens the softmax and emphasizes hard negatives. We tuned τ ∈ {0.07, 0.1}.

2.3 Input Representation: Log-Mel Spectrograms

We convert each clip to a log-mel spectrogram (e.g., 64 mel bins) and treat it as a single-channel image. This compresses dynamic range and aligns frequency resolution with human hearing. CNNs then learn local time–frequency patterns (harmonics, onsets, rhythms).

2.4 Architectures

• Baseline Encoder (3-Layer CNN): three conv blocks (Conv-BN-ReLU-MaxPool), channels 32→64→128, global pooling to a 128-D embedding plus a classification head. Proof-of-concept with moderate accuracy.
• Enhanced Encoder (4-Layer CNN): deeper conv stack, e.g., 64→128→256→512 with adaptive average pooling to a 256-D embedding. Increased capacity captures richer invariances; regularized with BN, augmentation, and early stopping.

2.5 Training Strategy & Tuning

Hybrid loss: contrastive + classification:

$$ L_{\mathrm{total}} = L_{\mathrm{InfoNCE}} + \alpha ,\bigl(L_{\mathrm{CE},\text{clip1}} + L_{\mathrm{CE},\text{clip2}}\bigr). $$

A small α keeps retrieval geometry primary while using supervised signals to stabilize features. Final best: LR (=10^{-4}), batch size (=32), embedding dim (=256), τ=0.07, α=0.25, Adam optimizer, up to 15 epochs with Reduce-on-Plateau scheduler.

Batch size: larger batches provide more in-batch negatives (better contrastive signal) but can hurt generalization. We used 32 due to hardware/memory.

2.6 Advanced Training Techniques

• Optimizers: Adam vs. AdamW (decoupled weight decay). In our setup, AdamW did not outperform Adam.
• LR scheduling: Reduce-on-Plateau and Cosine Annealing both helped over fixed LR.
• Augmentation:
  • SpecAugment (time/frequency masks, mild time warping) improved robustness.
  • Waveform augmentations (time-stretch, pitch-shift, noise/gain/compression) were less helpful overall in our settings.

---

3. State-of-the-Art with Pretrained Transformers

3.1 Large-Scale Pretrained Models

We evaluated Audio Spectrogram Transformer (AST) and Music Understanding Transformer (MERT)—self-attention encoders pretrained at scale, providing powerful audio/music representations.

3.2 Audio Spectrogram Transformer (AST)

Architecture: a ViT-style model over spectrogram patches. Self-attention captures long-range time–frequency dependencies.

Fine-tuning enhancements:

• Attention pooling: learn to weight frames by importance before forming a clip-level embedding.
• Learnable temperature τ for contrastive loss: lets training adapt similarity scaling.
• Multi-sample dropout: averages multiple dropout-perturbed heads to stabilize training.
• Label smoothing in the classification head for regularization.

3.3 Music Understanding Transformer (MERT)

Architecture & pretraining: transformer encoder trained with music-specific self-supervision (e.g., masked acoustic modeling with teacher signals), yielding strong melody/pitch/structure awareness.

Fine-tuning enhancements: the same set as AST (attention pooling, learnable τ, label smoothing). MERT proved especially robust under noise.

---

Part II: Experiments & Results

1. Dataset Construction

Using the Jamendo API we downloaded 1,000 tracks and, after filtering invalid items, retained 989 unique songs. We split 791 / 198 for train/validation (80/20). For generalization we collected another 200 disjoint tracks as a held-out test set. Audio is MP3; metadata (track id, title, artist, album, duration, genre) live in metadata.csv.

---

2. Classical Landmark Fingerprinting — Results

Grid-search over PNS, FAN, ETH (Top-1 Accuracy & Latency):

Show full grid (click to expand)

Label	PNS	FAN	ETH	Accuracy	Latency (s)
PNS=20_FAN=15_ETH=0.3	20	15	0.3	0.875632	0.593492
PNS=20_FAN=15_ETH=0.5	20	15	0.5	0.857432	0.593995
PNS=20_FAN=15_ETH=0.7	20	15	0.7	0.803842	0.589477
PNS=20_FAN=30_ETH=0.3	20	30	0.3	0.914055	0.603808
PNS=20_FAN=30_ETH=0.5	20	30	0.5	0.887765	0.608314
PNS=20_FAN=30_ETH=0.7	20	30	0.7	0.847321	0.595650
PNS=20_FAN=50_ETH=0.3	20	50	0.3	0.940344	0.637148
PNS=20_FAN=50_ETH=0.5	20	50	0.5	0.901921	0.623506
PNS=20_FAN=50_ETH=0.7	20	50	0.7	0.880688	0.618405
PNS=30_FAN=15_ETH=0.3	30	15	0.3	0.753286	0.618855
PNS=30_FAN=15_ETH=0.5	30	15	0.5	0.709808	0.632791
PNS=30_FAN=15_ETH=0.7	30	15	0.7	0.676441	0.641774
PNS=30_FAN=30_ETH=0.3	30	30	0.3	0.807887	0.647252
PNS=30_FAN=30_ETH=0.5	30	30	0.5	0.795753	0.589047
PNS=30_FAN=30_ETH=0.7	30	30	0.7	0.767442	0.585295
PNS=30_FAN=50_ETH=0.3	30	50	0.3	0.878665	0.593647
PNS=30_FAN=50_ETH=0.5	30	50	0.5	0.839232	0.581981
PNS=30_FAN=50_ETH=0.7	30	50	0.7	0.839232	0.581275
PNS=40_FAN=15_ETH=0.3	40	15	0.3	0.560162	0.570391
PNS=40_FAN=15_ETH=0.5	40	15	0.5	0.564206	0.568640
PNS=40_FAN=15_ETH=0.7	40	15	0.7	0.521739	0.566893
PNS=40_FAN=30_ETH=0.3	40	30	0.3	0.736097	0.569670
PNS=40_FAN=30_ETH=0.5	40	30	0.5	0.711830	0.574928
PNS=40_FAN=30_ETH=0.7	40	30	0.7	0.669363	0.569679
PNS=40_FAN=50_ETH=0.3	40	50	0.3	0.816987	0.572586
PNS=40_FAN=50_ETH=0.5	40	50	0.5	0.792720	0.573016
PNS=40_FAN=50_ETH=0.7	40	50	0.7	0.772497	0.572708

Best Top-1 accuracy (5s clean, 1,000 songs): 94.0% with (PNS=20, FAN=50, ETH=0.3).

Accuracy vs. clip length / noise (Top-1 & Top-5, 1,000 songs):

• Top-1:

SNR / Duration	2 s	5 s	10 s
20 dB	43.4%	89.3%	97.0%
10 dB	29.7%	83.3%	96.1%
0 dB	14.6%	62.1%	89.3%

• Top-5:

SNR / Duration	2 s	5 s	10 s
20 dB	57.2%	92.5%	98.4%
10 dB	41.7%	89.7%	97.7%
0 dB	23.6%	73.5%	94.2%

Accuracy vs. clip length / noise (Top-1 & Top-5, 200-song test set):

• Top-1:

SNR / Duration	2 s	5 s	10 s
20 dB	47.0%	95.5%	97.0%
10 dB	35.7%	85.5%	98.5%
0 dB	21.0%	67.0%	91.0%

• Top-5:

SNR / Duration	2 s	5 s	10 s
20 dB	63.0%	96.5%	99.0%
10 dB	50.5%	91.0%	98.5%
0 dB	34.0%	80.0%	95.5%

Plots (1,000-song tuning set):

Plots (200-song test set):

---

3. CNN Results

3.1 Baseline 3-Layer CNN

Default config: BATCH_SIZE=64, EPOCHS=15, LR=1e-3, TEMP=0.1, ALPHA=1.0, EMB_DIM=128. Results on the 200-song test:

• Top-1:

SNR / Duration	2 s	5 s	10 s
Clean	54.00%	59.00%	62.50%
20 dB	35.50%	48.50%	52.00%
10 dB	19.50%	31.50%	32.50%
0 dB	9.00%	8.50%	9.50%

• Top-5:

SNR / Duration	2 s	5 s	10 s
Clean	66.00%	67.50%	71.50%
20 dB	55.00%	59.50%	64.00%
10 dB	40.00%	42.00%	40.00%
0 dB	11.00%	14.00%	15.00%

3.2 4-Layer CNN — Grid Search (5 s, clean, validation)

Show full grid (click to expand)

LR	τ	α	Dim	Ep.	Batch	Val Loss	Val Acc (%)
0.0001	0.07	0.25	128	15	32	6.292580	64.04698
0.0001	0.07	0.25	128	15	64	7.335290	47.61074
0.0001	0.07	0.25	256	15	32	5.963996	71.42953
0.0001	0.07	0.25	256	15	64	7.286747	64.71812
0.0001	0.07	0.50	128	15	32	9.596525	67.40268
0.0001	0.07	0.50	128	15	64	10.842608	62.70470
0.0001	0.07	0.50	256	15	32	9.302552	70.08725
0.0001	0.07	0.50	256	15	64	10.854216	64.04698
0.0001	0.07	0.75	128	15	32	12.792898	65.38926
0.0001	0.07	0.75	128	15	64	14.163148	67.40268
0.0001	0.07	0.75	256	15	32	12.853767	66.06040
0.0001	0.07	0.75	256	15	64	14.019282	60.02013
0.0001	0.10	0.25	128	15	32	5.919953	69.41611
0.0001	0.10	0.25	128	15	64	7.348520	62.70470
0.0001	0.10	0.25	256	15	32	6.023814	59.34899
0.0001	0.10	0.25	256	15	64	7.345748	58.67785
0.0001	0.10	0.50	128	15	32	9.498577	59.73154
0.0001	0.10	0.50	128	15	64	10.707890	55.32215
0.0001	0.10	0.50	256	15	32	9.156086	71.10067
0.0001	0.10	0.50	256	15	64	10.928786	64.04698
0.0001	0.10	0.75	128	15	32	12.817372	69.41611
0.0001	0.10	0.75	128	15	64	14.366496	58.67785
0.0001	0.10	0.75	256	15	32	12.783500	70.75839
0.0001	0.10	0.75	256	15	64	14.362817	62.70470
0.0005	0.07	0.25	128	15	32	6.226688	59.34899
0.0005	0.07	0.25	128	15	64	6.547420	63.37584
0.0005	0.07	0.25	256	15	32	6.035327	68.74497
0.0005	0.07	0.25	256	15	64	6.702687	69.41611
0.0005	0.07	0.50	128	15	32	9.579094	55.32215
0.0005	0.07	0.50	128	15	64	10.484405	66.73154
0.0005	0.07	0.50	256	15	32	9.533671	68.07383
0.0005	0.07	0.50	256	15	64	10.100034	62.03356
0.0005	0.07	0.75	128	15	32	13.092276	56.66443
0.0005	0.07	0.75	128	15	64	13.816619	60.02013
0.0005	0.07	0.75	256	15	32	13.052927	60.69128
0.0005	0.07	0.75	256	15	64	13.750700	58.67785
0.0005	0.10	0.25	128	15	32	6.237953	59.34899
0.0005	0.10	0.25	128	15	64	6.803458	68.07383
0.0005	0.10	0.25	256	15	32	5.878330	66.06040
0.0005	0.10	0.25	256	15	64	6.936547	64.04698
0.0005	0.10	0.50	128	15	32	9.540131	64.71812
0.0005	0.10	0.50	128	15	64	10.201151	58.00671
0.0005	0.10	0.50	256	15	32	9.830219	69.41611
0.0005	0.10	0.50	256	15	64	10.392408	64.04698
0.0005	0.10	0.75	128	15	32	13.015230	64.04698
0.0005	0.10	0.75	128	15	64	13.481024	58.00671
0.0005	0.10	0.75	256	15	32	13.165791	68.07383
0.0005	0.10	0.75	256	15	64	13.586086	62.70470

3.3 Optimizer / Scheduler / Augmentation Grid

(fixed: batch=32, LR=1e-4, emb=256, temp=0.07, alpha=0.25)

Show table (click to expand)

Optimiser	Scheduler	Model	Augment.	Acc (%)	Time (min)
Adam	ReduceLROnPlateau	Encoder3Layer	none	60.02	81.66
Adam	ReduceLROnPlateau	Encoder3Layer	specaugment	62.53	94.88
Adam	ReduceLROnPlateau	Encoder3Layer	audioment_light	62.02	95.95
Adam	ReduceLROnPlateau	Encoder3Layer	audioment_heavy	66.06	96.77
Adam	ReduceLROnPlateau	Encoder4Layer	none	64.55	94.74
Adam	ReduceLROnPlateau	Encoder4Layer	specaugment	70.53	94.89
Adam	ReduceLROnPlateau	Encoder4Layer	audioment_light	69.49	95.94
Adam	ReduceLROnPlateau	Encoder4Layer	audioment_heavy	67.47	96.84
Adam	CosineAnnealingLR	Encoder3Layer	none	61.03	39.88
Adam	CosineAnnealingLR	Encoder3Layer	specaugment	64.55	94.93
Adam	CosineAnnealingLR	Encoder3Layer	audioment_light	66.97	95.93
Adam	CosineAnnealingLR	Encoder3Layer	audioment_heavy	67.07	96.73
Adam	CosineAnnealingLR	Encoder4Layer	none	63.06	94.88
Adam	CosineAnnealingLR	Encoder4Layer	specaugment	69.49	94.87
Adam	CosineAnnealingLR	Encoder4Layer	audioment_light	67.07	95.91
Adam	CosineAnnealingLR	Encoder4Layer	audioment_heavy	68.01	96.75
AdamW	ReduceLROnPlateau	Encoder3Layer	none	61.57	94.82
AdamW	ReduceLROnPlateau	Encoder3Layer	specaugment	62.53	94.83
AdamW	ReduceLROnPlateau	Encoder3Layer	audioment_light	62.02	95.85
AdamW	ReduceLROnPlateau	Encoder3Layer	audioment_heavy	66.06	96.77
AdamW	ReduceLROnPlateau	Encoder4Layer	none	65.04	94.98
AdamW	ReduceLROnPlateau	Encoder4Layer	specaugment	69.52	95.03
AdamW	ReduceLROnPlateau	Encoder4Layer	audioment_light	64.49	95.98
AdamW	ReduceLROnPlateau	Encoder4Layer	audioment_heavy	67.47	96.75
AdamW	CosineAnnealingLR	Encoder3Layer	none	63.54	94.84
AdamW	CosineAnnealingLR	Encoder3Layer	specaugment	64.49	94.89
AdamW	CosineAnnealingLR	Encoder3Layer	audioment_light	66.97	95.94
AdamW	CosineAnnealingLR	Encoder3Layer	audioment_heavy	67.07	96.79
AdamW	CosineAnnealingLR	Encoder4Layer	none	64.55	94.93
AdamW	CosineAnnealingLR	Encoder4Layer	specaugment	66.49	94.88
AdamW	CosineAnnealingLR	Encoder4Layer	audioment_light	68.01	95.96
AdamW	CosineAnnealingLR	Encoder4Layer	audioment_heavy	68.01	96.77

3.4 Enhanced CNN — Test Results (200 songs)

• Top-1:

SNR / Duration	2 s	5 s	10 s
Clean	56.50%	71.50%	73.50%
20 dB	45.00%	59.50%	57.50%
10 dB	31.50%	38.00%	43.00%
0 dB	14.50%	12.50%	15.00%

• Top-5:

SNR / Duration	2 s	5 s	10 s
Clean	71.50%	73.00%	82.00%
20 dB	62.00%	66.00%	72.00%
10 dB	48.00%	49.00%	52.00%
0 dB	21.00%	18.50%	23.00%

Embedding similarity distributions:

t-SNE of embeddings by genre:

---

4. Transformer Results

4.1 Audio Spectrogram Transformer (AST)

Baseline AST (200-song test):

• Top-1:

SNR / Duration	2 s	5 s	10 s
Clean	51.50%	78.50%	72.00%
20 dB	42.50%	73.50%	67.00%
10 dB	32.50%	71.50%	62.00%
0 dB	15.50%	50.00%	47.50%

• Top-5:

SNR / Duration	2 s	5 s	10 s
Clean	71.00%	89.00%	89.50%
20 dB	69.00%	86.00%	87.00%
10 dB	51.50%	86.00%	83.00%
0 dB	30.50%	73.00%	76.50%

Enhanced AST (attention pooling, learnable τ, multi-sample dropout, label smoothing):

• Top-1:

SNR / Duration	2 s	5 s	10 s
Clean	54.50%	74.50%	75.50%
20 dB	42.50%	75.00%	70.50%
10 dB	33.50%	71.00%	65.50%
0 dB	15.50%	50.50%	50.00%

• Top-5:

SNR / Duration	2 s	5 s	10 s
Clean	74.50%	89.00%	90.00%
20 dB	65.50%	89.00%	88.00%
10 dB	55.00%	80.50%	82.00%
0 dB	33.50%	66.00%	73.00%

Embedding similarity distributions:

t-SNE by genre:

4.2 Music Understanding Transformer (MERT)

Baseline MERT (200-song test):

• Top-1:

SNR / Duration	2 s	5 s	10 s
Clean	51.50%	84.50%	85.00%
20 dB	55.50%	77.50%	84.00%
10 dB	50.00%	69.50%	80.00%
0 dB	42.00%	61.50%	61.00%

• Top-5:

SNR / Duration	2 s	5 s	10 s
Clean	71.00%	91.00%	94.00%
20 dB	68.00%	91.00%	89.50%
10 dB	68.50%	82.50%	91.50%
0 dB	63.00%	79.50%	78.00%

Enhanced MERT (attention pooling, learnable τ, label smoothing):

• Top-1:

SNR / Duration	2 s	5 s	10 s
Clean	67.00%	84.00%	86.50%
20 dB	56.00%	75.50%	85.00%
10 dB	54.50%	69.50%	84.50%
0 dB	49.50%	61.50%	61.50%

• Top-5:

SNR / Duration	2 s	5 s	10 s
Clean	78.50%	92.00%	91.00%
20 dB	77.50%	91.50%	92.50%
10 dB	78.50%	87.00%	91.00%
0 dB	62.00%	77.00%	72.50%

Embedding similarity distributions:

t-SNE by genre:

---

5. Comparisons

• Fingerprinting vs. Learning-Based: Classical fingerprinting reaches ~89–94% Top-1 on clean 5-second clips, exceeding all CNNs and some transformer variants in low-noise conditions.
• CNN Baseline vs. Enhanced: 3-layer CNN achieves ~59% Top-1 (clean 5 s); 4-layer with tuning and SpecAugment reaches ~71% (+12 pp).
• AST vs. CNN: Baseline AST 78.5% Top-1 (clean 5 s) > Enhanced CNN 71.5%.
• AST Fine-tuning: With attention pooling etc., Top-1 on clean 5 s remains around the same or slightly lower—suggesting baseline AST is already strong.
• MERT vs. AST: MERT hits ~84.5% Top-1 (clean 5 s) and is more robust to heavy noise (0 dB: 42% vs. 15.5% for AST).
• Enhanced MERT: Advanced fine-tuning lifts clean 5 s Top-1 to ~86.5% and maintains strong performance across conditions.
• Top-5 Trends: Transformers (AST, MERT) yield >90% Top-5 on clean 5 s and degrade more gracefully with noise than CNNs.

Triplet analyses (200-song test, MERT embeddings):

• Most dissimilar triplets (lowest mean cosine):

Song A	Song B	Song C	Avg. Cosine
4883	5988	8832	0.6083
5988	6571	9201	0.6106
4883	5988	6297	0.6149
5988	6571	8832	0.6165
5988	6297	9201	0.6167

• Most similar triplets (highest mean cosine):

Song A	Song B	Song C	Avg. Cosine
7197	7202	7203	0.9737
5750	5751	5753	0.9659
1105	7197	7202	0.9640
220	225	3435	0.9626
5750	5752	5753	0.9615

Cosine-similarity heatmaps — most similar triplets:

Cosine-similarity heatmaps — most dissimilar triplets:

---

Conclusions

We compared a classical landmark-based audio fingerprinting pipeline against modern deep learning (CNN) and large pretrained transformer models (AST, MERT) for song identification:

• Classical fingerprinting remains outstanding on clean, short clips (Top-1 up to 94% at 5 s), fast and robust to partial matches.
• CNNs benefit from depth and hybrid losses but trail transformers; best enhanced CNN reached ~71% Top-1 on clean 5 s.
• Pretrained transformers shine: AST outperforms CNNs; MERT (music-specific SSL) is best overall, with ~84.5–86.5% Top-1 on clean 5 s and strong noise robustness.
• Generalization from tuning to held-out test is stable across methods.
• Embeddings exhibit clear structure (similarity distributions and t-SNE), supporting retrieval-style song ID.

Future directions: fuse classical fingerprints with learned embeddings, and explore larger or domain-specialized pretrained backbones for even stronger performance.