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

Abstract

Pitch shifting is a common voice editing technique in which the original pitch of a digital voice is raised or lowered. It is likely to be abused by the malicious attacker to conceal his/her true identity. Existing forensic detection methods are no longer effective for weakly pitch-shifted voice. In this paper, we proposed a convolutional neural network (CNN) to detect not only strongly pitch-shifted voice but also weakly pitch-shifted voice of which the shifting factor is less than ±4 semitones. Specifically, linear frequency cepstral coefficients (LFCC) computed from power spectrums are considered and their dynamic coefficients are extracted as the discriminative features. And the CNN model is carefully designed with particular attention to the input feature map, the activation function and the network topology. We evaluated the algorithm on voices from two datasets with three pitch shifting software. Extensive results show that the algorithm achieves high detection rates for both binary and multiple classifications.

A Parallel Classification Model for Marine Mammal Sounds Based on Multi-Dimensional Feature Extraction and Data Augmentation

“…Finally, according to the number of Mel filters selected, the Mel filter bank matrix H m ( k ) with size of 513 × 128 was generated and multiplied by the results of STFT to obtain a power spectrum of Mel scale, i.e., a 59 × 128 real matrix X m ( t ). The calculation process is shown in Equations (3) and (4) [ 63 ]: where X m ( t ) denotes the power spectrum matrix, L denotes the number of data frames in the direction of time axis, M denotes the number of Mel filter banks, S t ( k ) denotes the spectrum obtained by STFT, H m ( k ) is the expression of Mel filter banks, where o ( m ), c ( m ), and h ( m ) are frequency points, and their interval determines the type of filter banks.…”

Classification of Pitch and Gender of Speakers for Forensic Speaker Recognition from Disguised Voices Using Novel Features Learned by Deep Convolutional Neural Networks

“…Wu et al [20] considered electronic means of pitch disguise and classified the speech into disguised or original speech using MFCC static and correlation features and a novel classification algorithm using support vector machines (SVM). Identification of weakly pitch shifted voice is performed in view of the forensic scenario [16]. We have also considered electronic pitch disguise in our work and extended the problem addressed in the study [20] for the classification of disguised voice into high pitch or low pitch voices.…”

“…Identifying whether a given test speech is disguised or original is the first step in ASR from disguised voices. In some works, deep features and neural network classifiers are used for this classification [15][16][17][18]. This classification is done in literature using both prosodic and cepstral features [16,[18][19][20][21].…”

“…In some works, deep features and neural network classifiers are used for this classification [15][16][17][18]. This classification is done in literature using both prosodic and cepstral features [16,[18][19][20][21]. Specific types of disguises are considered in most of the works like pitch disguised voices [16,[18][19][20], creaky voices [9,17], mimicked voices [15,21] etc.…”

Data augmentation and deep neural networks for the classification of Pakistani racial speakers recognition

“…In another investigation, scientists employed an auditory DA strategy to achieve an 82.6 percent accuracy for Mandarin-English code flipping ( Long et al, 2020 ). As presented in Ye et al (2020) pitch shifting is frequently utilized in DA and achieved 90% accuracy. In addition, Damskägg & Välimäki (2017) employed the time-stretched data augmentation approach when performing DA-based fuzzy identification on various audio signals.…”