Article Preview
Top1. Introduction
Speech is the most natural way to communicate between humans. Nowadays, speech communication systems acquire, transmit, store, and process the information in digital form. However, the digital speech content can be imperceptibly altered by malicious, even amateur, users by using a variety of low-cost audio editing software. This creates a serious threat to the knowledge life cycle. Indeed, when hearing is no longer believing, the process of going from data to information, knowledge, understanding and, decision making is severely compromised (Farid, 2008). The consequences of this threat permeate a wide variety of fields, such as intellectual property, intelligence gathering, forensics, and news reporting to name a few. Currently, the methods to combat this threat in the field of digital speech forensics are still in their infancy. Therefore, there is an urgent need to advance the state-of-the-art in this field (Garcia-Romero & Espy-Wilson, 2010).
A first step to remedy the aforementioned threat is to extract forensic evidence about the mechanism involved in the generation of the speech recording by analyzing only the speech signal (Garcia-Romero & Espy-Wilson, 2010). That is, to identify the acquisition device by assuming that the devices along with their associated signal processing chain leave behind intrinsic traces in the speech signal. Indeed, the electronic devices, especially when including a microphone, cannot have exactly the same frequency response due to tolerances in the production of their electronic components and the different designs employed by the various manufacturers (Hanilci, 2012). This implies that the recorded speech can be considered as a signal whose spectrum is the product of the genuine speech spectrum, driving the acquisition device, and the frequency response of the latter. Consequently, the recorded speech signal can be exploited in device identification, following a blind-passive approach, as opposed to active embedding of watermarks or having access to input-output pairs (Garcia-Romero & Espy-Wilson, 2010).
Although there are significant advances in image forensics (Farid, 2008), audio forensics are less developed (Maher, 2009). Few exceptions include the authentication of MP3 (Yang, 2008) and the authentication of speakers’ environment (Oermann, 2005; Kraetzer, 2007; Malik & Farid, 2010). Similarly, a few automatic acquisition device identification systems have been developed. For instance, a method for the classification of 4 microphones has been proposed in (Kraetzer, 2007). The speech signal is parameterized by employing time domain features and the mel-frequency cepstral coefficients (MFCCs). The identification of the microphones is performed by a Naive Bayes classifier at a short-time frame level. Accuracies on the order of 60-75% have been reported. In (Garcia-Romero & Espy-Wilson, 2010), the identification of 8 landline telephone handsets and 8 microphones is addressed. In particular, the intrinsic characteristics of the device are captured by a template constructed by appending together the means of a Gaussian mixture which have been trained using linear and mel-scaled cepstral coefficients extracted by speech recordings of each device. Classification accuracies higher than 90% have been achieved, when a support vector machine (SVM) classifier was used. Recently, a robust system for the identification of cell-phones has been proposed in (Hanilci, 2012). In particular, when the MFCCs, extracted from speech recordings, are classified by an SVM, 14 different cell-phones are identified with an accuracy of 96.42%.