Detecting Discontinuities in Audio Recordings

Every audio recording is made in a specific acoustic environment, on specific equipment, at a specific moment in time. These conditions produce a set of physical properties that are consistent throughout the recording: the spectral colour and level of the background noise, the characteristics of room reverberation, the phase relationships between signal components, and any periodic signals from mains power or mechanical sources. When two segments from different recording sessions are joined, the join point is a discontinuity in one or more of these properties.

The most visible signatures are energy transients at the cut boundary. A hard cut, where one segment ends and another begins at a sample boundary, creates a sudden jump in waveform amplitude that produces a click or pop in the audio and an impulsive spike in the spectrogram. A cross-fade edit is harder to detect visually but still leaves a transition zone where the spectral properties of both segments overlap in an unnatural mixing pattern. Re-recording, where a segment is played back and a new segment is recorded over it, leaves artefacts from the re-recording chain: double microphone convolution, possible generation loss in analog systems, and spectral colouration from the playback speaker and re-recording room.

Background noise is the most consistent discriminator between segments recorded in different sessions. Even in the same room, the noise floor will differ if the air conditioning cycles, a window is opened, or background traffic changes. Across different rooms or buildings, the spectral profile of the noise will differ substantially. Automated noise profiling algorithms compute a statistical model of the noise in each quiet interval and flag points where the noise model changes abruptly. These changes are not conclusive evidence of editing, since natural acoustic events can cause similar shifts, but they define the candidate discontinuity locations for detailed examination.

The spectrogram is the analyst's primary visual display. A short-time Fourier transform (STFT) converts successive overlapping windows of the audio signal into a matrix of frequency bins versus time frames, with amplitude encoded in colour. The choice of window length determines the resolution tradeoff: a long window gives fine frequency resolution but poor time resolution, a short window gives fine time resolution but smears frequency detail. Forensic audio practice typically uses two or three spectrograms at different window lengths to capture both fine temporal events and fine spectral features.

On a well-parameterised spectrogram, discontinuities appear as abrupt vertical changes: a sudden shift in the level of a narrowband background tone such as 50 Hz hum or HVAC resonance, a vertical stripe of elevated broadband energy at the cut point (the click transient), or an abrupt change in the spectral slope of the noise floor. Room-tone truncation appears as a sharp vertical edge in the decay region of a spectrogram frame where the reverberant energy should continue for another 50 to 200 milliseconds.

Narrowband tonal components are particularly useful for detecting discontinuities because they must maintain phase continuity across any edit point if the recording is unmodified. Mains hum at 50 or 60 Hz and its harmonics, fluorescent-light flicker, and mechanical resonances all contribute tonal components to a recording. An examiner who tracks the instantaneous phase of a 50 Hz component across the recording and finds an unexplained phase jump at a candidate discontinuity has strong corroborating evidence of an edit at that point.

Phase-discontinuity analysis operates directly on the audio waveform. The instantaneous phase of a narrowband-filtered version of the signal is extracted using the Hilbert transform or a phase-locked loop. In a continuous recording, the phase of any periodic component progresses monotonically. An edit that splices two segments together at an arbitrary point will almost certainly introduce a phase jump, because the periodic component in the second segment will not be at the phase value predicted by the continuation of the first segment.

Energy transient analysis uses short-time energy functions, typically the squared envelope of the signal, to locate abrupt amplitude events. In a continuous speech or environmental recording, the energy envelope changes at speech onsets and offsets, breath sounds, and acoustic events in the environment, all of which have characteristic gradients. A vertical jump in the energy envelope that is not associated with a speech event or an acoustic event in the scene is a candidate edit point. Many automated discontinuity detectors use this energy gradient as a first-pass filter to locate candidate points for detailed spectrographic review.

A well-established automated approach is the Bayesian change-point detector applied to spectral features. The algorithm models the spectral properties of the recording as a piecewise stationary process and uses Bayes factors to identify time points where the spectral model must change. Change points with high Bayes factors are candidate discontinuities. This approach was validated in peer-reviewed literature by Gallagher et al. (2011) and has been applied in several court-admissible examinations in the United States and the United Kingdom.

Background noise profiling extracts the ambient noise signal from quiet intervals in a recording and builds a statistical model of its properties: mean level in decibels, spectral slope, dominant tonal components, and statistical stationarity. The model is computed independently for each quiet interval and the models are compared across the recording. A recording made in a single continuous session in a stable acoustic environment will produce statistically similar noise models throughout. A recording assembled from segments recorded at different times or in different locations will contain noise models that differ in some measurable respect.

Room-tone analysis focuses specifically on the reverberation characteristics of the space. Each enclosed room has a set of resonant modes (room modes) at frequencies determined by the room's dimensions, and a characteristic reverberation time (RT60, the time for the sound to decay by 60 dB). These properties can be estimated from the recording by analysing the decay rate and spectral shape of the reverberant tail that follows each speech event or acoustic event in the scene. If different parts of a recording exhibit significantly different RT60 values or different room-mode signatures, the segments were recorded in different acoustic environments.

A practical limitation is that room properties can change naturally within a single recording session: a participant moves position, a door opens or closes, a soft furnishing is introduced. The examiner must assess whether a measured change in room acoustics is consistent with a natural acoustic event that is audible in the recording, or whether it is abrupt and unexplained. Natural acoustic transitions are gradual; splice-point acoustic transitions are instantaneous.

The ENF method provides a uniquely powerful check on the temporal integrity of a recording because it links the audio content to an independent external reference. Recordings made on mains-powered devices, or battery-powered devices operating near mains infrastructure, capture the fluctuating mains frequency as a tonal component at 50 Hz or 60 Hz (and often at harmonics). The fluctuation pattern is caused by variations in power generation and grid load and is essentially unique to a specific geographic grid zone and time window.

National grid operators in the United Kingdom, the United States, Germany, India, and several other countries have maintained historical ENF reference databases. The UK National Grid has provided reference data used in British court proceedings. The examiner extracts the instantaneous ENF from the questioned recording and correlates it against the reference database to determine the time window during which the recording was made. If the ENF pattern in the questioned file is discontinuous or if it correlates with two non-contiguous time windows, the recording contains segments from different sessions, confirming editing.

ENF evidence has been used in courts in the United Kingdom, Spain, and the United States. In India, section 79A of the Information Technology Act 2000 designates the Indian Computer Emergency Response Team as the authority on electronic evidence standards, and examiners working under the Bharatiya Sakshya Adhiniyam 2023 present ENF findings as part of a broader authentication report. The European Union's Digital Services Act framework and the eIDAS regulation create an ecosystem in which provenance-based authentication, including ENF, is increasingly formalised.

The authentication report for audio discontinuity analysis must document four things: the provenance of the recording (how it was acquired, by whom, under what chain of custody), the analytical methods applied and the instrumentation or software used, the specific findings at each candidate discontinuity, and the examiner's conclusion expressed in terms of a likelihood or probability scale. Binary conclusions (authentic or not authentic) are not appropriate because the analysis is probabilistic: some findings are strong indicators, others are equivocal, and the absence of detected discontinuities does not prove the recording is unmodified.

The conclusion scale commonly used in forensic audio follows the SWGDE (Scientific Working Group on Digital Evidence) verbal probability scale, similar to the scale used in fingerprint and DNA reporting: from 'the findings are highly consistent with authentic, unedited audio' through intermediate levels to 'the findings are highly consistent with editing or manipulation'. Courts in the United States operating under the Daubert standard will assess the examiner's methodology against criteria of testability, peer review, known error rate, and general acceptance. UK courts apply the Bonython criteria. These standards require the examiner to cite the validation literature for each method applied.

Exhibit materials for court presentation typically include annotated spectrograms with candidate discontinuity points marked, waveform plots showing energy transients, and noise-profile comparison graphs. ENF correlation plots, showing the questioned ENF trace overlaid on the reference database, are particularly compelling visual exhibits because they translate a technical finding into a readily understood time-alignment display. The examiner must be prepared to explain each exhibit in lay terms and to respond to cross-examination on the limitations and error rates of each method.

Chain of custody for the audio file itself is a prerequisite for any authentication report. An examiner who cannot confirm that the file presented for analysis is the same file recovered from the original device or storage medium cannot rule out the possibility that a copy was substituted. Hash verification (SHA-256 or equivalent) at each transfer point, documented in a continuity log, satisfies this requirement. The relevant standard in India is the Bharatiya Sakshya Adhiniyam 2023, which sets out conditions for electronic evidence admissibility. Equivalent requirements apply under the US Federal Rules of Evidence Rule 901(b)(9) and the UK's Police and Criminal Evidence Act 1984 codes of practice. See also Chain of Custody for Digital Media.