Karlheinz Brandenburg, Fraunhofer Institute for Digital Media Technology, Ilmenau, Germany
- "Lossy compression in digital evidence: Easier or more difficult?"
Abstract: The talk connects the topics of audio coding (via mp3 or similar algorithms) to forensics. If an audio signal has been stored in a low bit-rate format, finding forensic evidence depends on several circumstances.
Lossy compression can add additional evidence, like information about the codecs used, whether tandem coding has been employed, or whether all the audio has been encoded in one step or assembled from different sources.
On the other hand, audio coding might make forensic tasks like speaker identification or verification more difficult. Forensic traces could become more weak or in some cases completely cancelled because of audio coding.
The talk will give a short introduction to audio coding showing which steps make the analysis more difficult and which additional traces are there due to the compression.
In total, a signal which has been subject to lossy compression needs to be analyzed with care, but it does not necessarily show less possibilities for forensic analysis.
Udo Zölzer, Helmut Schmidt University and University of the Federal Armed Forces Hamburg, Germany
- "Source-Filter Processing for Audio Forensics"
Abstract: Source-filter processing is a signal processing technique for modeling of an audio signal in the time and frequency domain. The modeling is performed by the extraction of an excitation signal or source signal and a time-varying filter or spectral envelope from the input audio signal. Basic signal processing techniques will be introduced for performing this source-filter separation. The use of this decomposition for modifying the excitation signal and the spectral envelope for specific tasks of audio forensics will be discussed.
Catalin Grigoras, National Center for Media Forensics, University of Colorado Denver, USA
- "Forensic Audio Authentication"
Abstract: The goal of this tutorial is to provide a summary review of the latest developments in conducting comprehensive examinations of digital audio authenticity which rely on the results of multiple analyses to inform an ultimate scientific finding or unbiased opinion. Digital audio authentication is a complex process of establishing the provenance of a questioned recording to determine whether it is consistent with an original one or if there is evidence of editing. This tutorial presents the organization of several techniques in a logical manner for the authentication of digital audio recordings. Special attention has been given to interpreting results from individual analyses and incorporating them into a holistic view of a recording’s authenticity where a finding can be corroborated against the results of other analyses. Only in this way can an examiner present a conclusion with confidence and assurance that all possible hypotheses have been exhausted in the execution of this important endeavor. The framework for digital audio authentication that will be discussed involves accurate, repeatable, reliable, unbiased, and scientific analyses derived from peer reviewed publications in order to meet court guidelines or case precedence, best practice recommendations, and the appropriate criteria for international legal systems.
Eliud Bonilla, Johns Hopkins University Applied Physics Laboratory, Laurel, MD, USA
- "Automatic Speaker Recognition"
Abstract: This tutorial will present an introductory overview of forensic automated speaker recognition analysis. Fundamental concepts, principles, and methodology will be explained from a practitioner’s perspective. There will be both a lecture segment as well as live demonstrations. By the end of this tutorial attendees will have a basic understanding of the underlying concepts and principals of forensic automated speaker recognition. The live demonstrations will provide an appreciation of the discipline, best practices, and methods for analysis.
Andy Sarroff, iZotope, Cambridge, MA, USA
- "Deep Learning for Audio Synthesis, Separation, and Enhancement"
Abstract: Audio presents challenges to machine learning that are unique from other domains, such as images and financial data. At typical sampling rates for speech, the digital audio waveform is rather high dimensional. Audio's time varying properties induce a version of the image correspondence problem, where waveforms or spectrograms that are very distant to each other in, say, Euclidean space, correspond to perceptually similar sounds. Audio has local and global dependencies covering time scales of multiple orders of magnitude, such as local structure at the sample level, phonemic structure for speech at the level of milliseconds, and semantic structure existing across seconds and minutes. Human perception of audio quality may be sensitive to small perturbations to signal integrity, yet differentiable training objectives don't always correspond to subjective experience. Nevertheless amazing advances have been made in deep learning for audio, including the areas of enhancement, separation, and synthesis. This tutorial will provide a high level overview of the building blocks needed for deep learning of audio, as well as a discussion of the state of the art.
Durand Begault, Audio Forensic Center, San Francisco, CA, USA
Robert Maher, Montana State University, Bozeman, MT, USA
Steven Beck, Beck Audio Forensics, Austin, TX, USA
- "Gunshot Analysis"
Abstract: Recordings of gunfire present multiple challenges to audio forensic practitioners. This workshop includes descriptions of case experiences from three experts in the field. Topics include simultaneous recordings of the same gunshot sequence from different vantage points; earwitness testimony; signal transformation by the recording process; and gunshot detection systems. Questions and observations from the audience are encouraged.