SFXlearner

This is the codebase for paper "Automatic Recognition of Cascaded Guitar Effects" at DAFX23.

A bunch of rendered examples can be found here.

Please refer to the colab notebook for a run through of the workflow.

Effects list

Number	Effect name	pySoX Function name	parameters
1	overdrive	overdrive()	{'gain_db': 5}
2	distortion	overdrive()	{'gain_db': 15}
3	chorus	chorus()	{'n_voices': 5}
4	flanger	flanger()	{'depth': 5, 'phase': 50}
5	phaser	phaser()	{}
6	tremolo	tremolo()	{}
7	reverb	reverb()	{'reverberance': 80}
8	feedback_delay	echos()	{'n_echos': 3, 'delays': [200,400,600], 'decays':[0.4,0.2,0.1], 'gain_out':0.5}
9	slapback_delay	echo()	{'n_echos': 3, 'delays': [200,400,600], 'decays':[0.4,0.2,0.1], 'gain_out':0.5}
10	low_boost	bass()	{'frequency': 200, 'gain_db': 10}
11	low_reduct	bass()	{'frequency': 200, 'gain_db': -10}
12	hi_boost'	treble()	{'frequency': 8000, 'gain_db': 20}
13	hi_reduct'	treble()	{'frequency': 8000, 'gain_db': -20}

SingleFX results

Dataset type	n_classes	Feature Type	Using Clean	Epochs	Validation Accuracy (Highest)	Notes
1on1	13	MFCC mean	No	100	0.41
1on1	13	MFCC mean	Yes	100	0.85
1onN	13	MFCC mean	No	100	0.55	converge fast
1onN	13	MFCC mean	Yes	100	0.96	converge fast

MultiFX results

Model type	Test set	With clean	effect method	n_classes	Feature Type	micro F1	macro F1	Notes
CRNN	guitarset test split	Yes	[1,5]	13	Mel Spectrogram	0.999	0.999	converge slow
resnet18	guitarset test split	Yes	[1,5]	13	Mel Spectrogram	0.999	0.999
baseline	guitarset test split	Yes	[1,5]	13	MFCC	0.951	0.952	MFCC+MLP
sampleCNN	guitarset test split	Yes	[1,5]	13	Raw audio	0.864	0.684
CRNN	IDMT-SMT-GUITAR	Yes	[1,5]	13	Mel Spectrogram	0.963	0.961	converge slow
resnet18	IDMT-SMT-GUITAR	Yes	[1,5]	13	Mel Spectrogram	0.968	0.970
resnet14	IDMT-SMT-GUITAR	Yes	[1,5]	13	Mel Spectrogram	0.963	0.955
resnet10	IDMT-SMT-GUITAR	Yes	[1,5]	13	Mel Spectrogram	0.958	0.950
resnet6	IDMT-SMT-GUITAR	Yes	[1,5]	13	Mel Spectrogram	0.926	0.917
baseline	IDMT-SMT-GUITAR	Yes	[1,5]	13	MFCC	0.779	0.772	MFCC+MLP
sampleCNN	IDMT-SMT-GUITAR	Yes	[1,5]	13	Raw audio	0.804	0.640
CRNN	guitarset test split	No	[1,5]	13	Mel Spectrogram	0.967	0.968	converge slow
resnet18	guitarset test split	No	[1,5]	13	Mel Spectrogram	0.958	0.965
baseline	guitarset test split	No	1,5]	13	MFCC	0.892	0.897	MFCC+MLP
sampleCNN	guitarset test split	No	[1,5]	13	Raw audio	0.877	0.778
CRNN	IDMT-SMT-GUITAR	No	[1,5]	13	Mel Spectrogram	0.856	0.851	converge slow
resnet18	IDMT-SMT-GUITAR	No	[1,5]	13	Mel Spectrogram	0.876	0.906
resnet14	IDMT-SMT-GUITAR	No	[1,5]	13	Mel Spectrogram	0.848	0.832
resnet10	IDMT-SMT-GUITAR	No	[1,5]	13	Mel Spectrogram	0.860	0.844
resnet6	IDMT-SMT-GUITAR	No	[1,5]	13	Mel Spectrogram	0.830	0.811
baseline	IDMT-SMT-GUITAR	No	[1,5]	13	MFCC	0.704	0.696	MFCC+MLP
sampleCNN	IDMT-SMT-GUITAR	No	[1,5]	13	Raw audio	0.697	0.623