Identification of Weakly Pitch-Shifted Voice Based on Convolutional Neural Network

Table 2

Detection performance of strongly pitch-shifted voice in binary classification.

Pitch shifting software	Training dataset	Testing dataset	Detecting method
			[6] LFCC + GMM		[8] MFCC + GMM		Proposed
			Rate	FAR	Rate	FAR	Rate	FAR
Audition	TIMIT	TIMIT	99.86	0.02	99.88	0.02	99.54	0.10
	TIMIT	UME	97.60	1.10	98.06	1.19	95.89	1.52
	UME	TIMIT	99.52	0.36	98.58	0.02	97.51	1.45
	UME	UME	99.79	0.15	99.79	0.12	99.49	0.12
GoldWave	TIMIT	TIMIT	99.97	0.00	99.94	0.01	99.58	0.05
	TIMIT	UME	97.93	0.75	96.82	2.04	96.29	1.53
	UME	TIMIT	99.72	0.05	98.45	0.01	98.44	1.17
	UME	UME	99.87	0.02	99.70	0.07	99.12	0.36
Audacity	TIMIT	TIMIT	99.98	0.00	99.97	0.00	99.97	0.00
	TIMIT	UME	99.13	0.44	97.57	2.10	99.78	0.07
	UME	TIMIT	99.97	0.01	98.72	0.00	99.96	0.01
	UME	UME	99.97	0.00	99.95	0.00	99.84	0.11

Bold values represent the best performance under same circumstances (in same row) of the three methods. For criteria detection rate (Rate), higher is better. For criteria false alarm rate (FAR), lower is better.