Audio Forensics - Free Online Audio Analysis Tool

Load an audio file to read its encoder chain.

An encoder tag records what the encoding software wrote, not what is true - it can be edited, copied from another file, or left behind by a tool that merely rewrapped the audio. What is hard to fake is consistency: frame headers that change part-way through the file, a lowpass filter far below what the claimed bitrate allows, or a source sample rate that disagrees with the frames actually written all indicate the audio passed through more than one encoder.

Noise floor: --

Segments found: --

Largest floor change: --

Too-clean silence: --

The forensic signal is a quiet segment whose floor sits far below the surrounding floor, or an abrupt floor change across a boundary - not the mere existence of quiet passages. Click a segment to hear it.

Real silence still contains irregular microphone and converter noise, so a quiet segment whose level barely varies, whose samples are almost all exactly zero, or whose amplitude distribution is unusually narrow is flagged as unusually uniform. That is an indicator only - noise gates, muted intros and lossy codecs with silence detection all produce the same statistics with nothing concealed.

Load an audio file to run this pass.

Discontinuity score over time

Every score here is measured against this recording's own statistics, not an absolute threshold - the same join looks entirely different in a whisper and in a drum track. A candidate that arrives together with a large level change is reported as a content change rather than a join, because that is what an ordinary hard cut or a closing noise gate produces. None of this proves an edit: crossfades hide joins completely, and clean edits in quiet material can leave nothing measurable at all.

Brighter regions carry more energy at that time and frequency; use the Spectrogram Floor control in the display rail to adjust sensitivity.

This is a controlled bit-depth and bandwidth stress test - a deliberate proxy for lossy-codec sensitivity, not a simulation of any specific real encoder. Treat differences as a comparative indicator only, not proof of editing.

Idle.

Absence of hum proves nothing, and a smooth trace does not authenticate a recording. Without a reference grid-frequency database for the relevant region and period, this tool can assess internal consistency only - it cannot date or geolocate a recording.

Load a stereo file to run this pass.

Left minus right difference

Correlation over time

A correlation near +1 means the channels are near-identical, near 0 that they share little, and negative values that they are partly polarity-inverted. All three occur legitimately - wide reverb, mid/side processing and deliberately decorrelated effects all lower correlation without anything having been edited. Read a sudden change against its neighbours, not the absolute number.

Idle.

Each keypad digit is two simultaneous tones, one from a low group and one from a high group. A block is only accepted when the two tones account for most of its energy, each clearly beats every other tone in its own group, their power ratio sits inside the range real telephone equipment produces, neither carries a strong second harmonic, and the pair lasts at least 40 ms. Those checks exist because speech would otherwise decode as a stream of digits - and they can still be defeated by heavy noise, aggressive compression or a narrow-band codec, which will cause genuine digits to be missed rather than invented.

Idle.

This is a heuristic indicator only, never a verdict. Detecting synthetic speech is an unsolved problem: many legitimate recordings will show some of these patterns, and many synthetic ones will show none.