Contents
A spectrogram is a picture of a recording’s frequency content over time, and it is the primary instrument an audio examiner reads. Louder frequencies show brighter, and patterns that are inaudible in playback become visible as shapes. Four kinds of forensic trace surface on the same image: the mains-hum line that dates a clip, the discontinuities a splice leaves, the hard high-frequency cutoff a low-bitrate codec imposes, and the background events that place a recording. None of these is a verdict, and one caveat governs all of them: “evidence of discontinuities is not evidence of editing” (ENFSI).
What does a spectrogram show?
A spectrogram plots time along one axis, frequency along the other, and intensity as brightness, computed by a short-time Fourier transform that slices the recording into overlapping windows. The result turns structure the ear glosses over into something an examiner can point at: the harmonic stacks of a voice, the broadband smear of noise, the clean bands of a tone. It is not an analysis in itself, it is the lens through which the other analyses are done, which is why it is the first thing an examiner opens.
How is the ENF hum read?
The clearest forensic feature is often the mains hum, a faint horizontal line sitting at 50 or 60 hertz with fainter harmonics above it. Its tiny wobble over time is the Electric Network Frequency, and tracking it is the basis for dating a recording against a grid archive, sometimes to the second. The extraction follows a statutory protocol of DC removal, a narrow bandpass around the line, decimation and spectral analysis (ENFSI, 2009). A break or a sudden phase jump in that line is also one of the surer signs of an edit. The method is covered in ENF analysis.
What do splice seams look like?
An edit tends to leave a mark a trained eye can catch on the spectrogram: a vertical broadband click where two segments were butted together, an abrupt change in the noise floor, a discontinuity in the hum line, or a shift in the background bed. These are read as continuity problems, points where the recording stops being internally consistent. Every one is a supporting signal rather than a conclusion, because impulsive noises and ordinary edits can look alike, which is why the reading is confirmed by ear and cross-checked. The full set of methods is in audio tampering detection.
What does compression leave on the picture?
Lossy compression writes itself across the top of the spectrogram as a hard horizontal ceiling: a low-bitrate codec discards the highest frequencies, leaving a sharp cutoff where the energy simply stops, often with a shelf whose height hints at the rough bitrate. That ceiling is a strong tell that a file presented as pristine has in fact been through a lossy codec, and whether a recording was ever lossily compressed is detectable at about 98.6 percent independent of the codec (Hennequin, Royo-Letelier and Moussallam, 2017). Reading the cutoff is the visual entry point to double-encoding detection.
Can a spectrogram place a recording?
Background acoustics on the spectrogram place a recording to a class, never an address. Reverberation reveals the room: Moore, Brookes and Naylor (2013) identified the room class at 96.1 percent across 22 rooms, and Baum and colleagues (2022) estimated a blind roomprint from speech at up to 93.6 percent, though both collapse when the microphone moves. Naming a specific background sound is weaker, in the range of 30 to 44 percent, and general scene category sits near 60 percent across ten classes. The ceiling is a grid region plus a room class plus a setting, never a street.
How far does the picture go?
The spectrogram shows only what survived to be shown. A re-encode erases the codec history above its own cutoff, a re-recording resets the whole chain, and a trace too faint to have been captured is simply not there. Each pattern the examiner reads is reported as a strength of support for a proposition, not a verdict (ENFSI, 2015), and earns weight only when independent traces agree. It is the shared lens behind every read in what forensics can learn from a recording.
Sources
- Hennequin, Royo-Letelier, Moussallam (2017). Codec Independent Lossy Audio Compression Detection. IEEE ICASSP 2017.
- Moore, Brookes, Naylor (2013). Roomprints for Forensic Audio Applications. IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA) 2013.
- Baum, Cuccovillo, Yaroshchuk, Aichroth (2022). Environment Classification via Blind Roomprints Estimation. arXiv:2209.07196.
- European Network of Forensic Science Institutes (2009). Best Practice Guidelines for ENF Analysis in Forensic Authentication of Digital Evidence (FSAAWG-BPM-ENF-001).
- European Network of Forensic Science Institutes (2015). ENFSI Guideline for Evaluative Reporting in Forensic Science (STEOFRAE).