Soundraw music detection

How Soundraw assembles a track

Soundraw is aimed at video creators, podcasters and agencies who need a licensed instrumental bed rather than a song. You pick genre, mood, length and energy, and the service returns an arrangement made of discrete sections — intro, build, drop, outro — that you can swap, shorten or mute individually before export.

That editable-section model is the defining property. Where a full-song generator commits to one continuous pass, Soundraw output is explicitly modular: sections are internally consistent by design, because they are meant to be interchangeable. The result is a track whose parts line up to the grid and to each other far more exactly than a played arrangement would.

There is also usually no generated lead vocal. Several of the cues that make vocal-led generations easy to question — breath shape, mispronounced stress, identical phrasing between repeats — simply are not present in the file, so the reading leans entirely on instrumental structure.

What the analysis actually measures

The engine has no idea what Soundraw is and never names it. It measures the file and reports how ordinary those measurements look next to human production. On instrumental library-style material, these are the readings that tend to carry information.

• Onset timing against an inferred grid — section-assembled music places hits with near-zero deviation, where players drift by milliseconds.
• Self-similarity across the track — repeated sections that are acoustically near-identical rather than merely similar.
• Spectral ceiling and high-band density — a typical delivery is bright and evenly filled to the codec limit.
• Stereo-field movement over time — a wide but static image, with side-channel energy that barely changes between sections.
• Cross-segment tonal agreement — the intro and the outro measuring almost the same, which is unusual in a performance.

Frequency distribution and production style

Soundraw output is produced to sit under speech. That intent shapes the spectrum in ways you can see before you read any probability: a controlled low end that does not wander, a mid-range that is deliberately uncluttered so a voice-over can sit on top, and a bright, consistently filled top end that gives the bed presence at low playback level.

Section boundaries in this material are often unusually clean — energy changes across the whole spectrum at once, on the bar, with no overlap. Human arrangements almost always smear those boundaries, because a cymbal rings over the barline and a player anticipates the downbeat.

None of this is a signature. It is the sound of production-library music, and a competent human composer writing to the same brief will produce a similar profile. That overlap is precisely why we weight structure and timing alongside spectrum instead of calling brightness alone a verdict.

What to listen for before you run anything

Listen once end to end with headphones. None of these is proof on its own, but several together make a track worth analysing carefully rather than waving through.

• Sections that begin and end exactly on the bar, with no fill, pickup or ring-over.
• A repeated section that sounds not just similar but acoustically identical to its earlier appearance.
• Energy that steps up in tiers rather than building continuously.
• A drum performance with no variation in velocity between equivalent hits.
• A stereo image that never moves, however much the arrangement changes underneath it.

What survives editing and re-export

Most Soundraw tracks in circulation have been cut to picture, faded, ducked under a voice-over or re-encoded by a video platform. Each step overwrites part of what the engine reads.

• Ducking and volume automation break loudness and crest-factor evidence immediately.
• Cutting a track to length removes repeated sections, which weakens the self-similarity reading.
• Platform re-encoding shifts spectral-ceiling evidence and lowers confidence in the high band.
• Grid-exact timing is the most durable of the group — it survives level changes and moderate re-encoding.
• Exported stems mixed by hand in a DAW can land in the inconclusive band, which is the honest answer for a genuinely hybrid file.

Where a Soundraw check goes wrong

Programmed human music is the problem case. A producer working in a DAW with quantised MIDI, copied sections and a loudness-maximised master produces measurements that overlap heavily with section-assembled generation. Electronic, lo-fi and library genres are the worst affected, and our electronic-music study documents that gap rather than hiding it.

The failure runs the other way too. A short Soundraw bed pulled from a compressed video, sitting at low level under dialogue, offers the engine very little to work with. The correct output there is inconclusive, and the report says inconclusive instead of guessing.

As always, there is no attribution. A raised estimate means the file measures like machine-generated music; it never means the file came from Soundraw, and the result screen will not say otherwise.

Why this page exists

Semrush records roughly 590 US searches a month for “soundraw ai”, and a share of that interest is people who already have a track in hand and want to know how a detector will read it — creators checking a licensed bed before publishing, and reviewers checking submissions. This page exists to answer that specific question honestly rather than to sell an attribution capability that does not exist.

Soundraw detection FAQ