Read the verdict in layers

Begin with the named outcome

The primary result is the provider's category for the file: AI-generated vocals, AI-generated instrumental, no signs of AI generation, or inconclusive. The application maps that category for display; it does not create a verdict by applying its own percentage threshold.

AI-generated vocals

The provider associated the primary AI signal with vocals. This does not identify a singer, prove cloning or separate the vocal stem.

AI-generated instrumental

The provider associated the primary AI signal with the non-vocal music. It does not identify a generator or establish which production stage involved AI.

No signs of AI generation

The provider returned its not-AI category for this version of the file. That is not a certificate of human authorship; editing and unfamiliar generators can evade classification.

Inconclusive

The provider did not supply a category that could be reported safely. Keep this outcome as uncertainty rather than converting it into either side.

Read the percentages as component indicators

Vocal and instrumental percentages come directly from the specialist service. They are displayed separately because they describe different components of its assessment. This site performs no stem separation and calculates no blended overall AI percentage.

Confidence has three bands

Low, Medium and High describe the provider's confidence in a component reading. A confidence label does not turn a probability into proof. Review it with the verdict, timeline, source-file quality and production history.

Use the timeline to locate disagreement

Where the provider returns window-level values, the timeline shows how its readings vary through the recording. A changing timeline can help you choose sections for review, but it cannot tell you why they differ. Arrangement changes, edits, compression and hybrid production are all plausible causes.

Quality changes the question

• Lossy copies: MP3 and AAC encoding discard signal detail.
• Short excerpts: five seconds is accepted, but a minimum is not a promise of useful evidence.
• Heavy production: limiting, remixing and re-recording can alter cues.
• Hybrid tracks: human and generated elements can coexist in one mix.

What the report cannot establish

• Who authored, owns or performed the work
• Which generator or model was used
• Whether a platform rule or law was violated
• Whether a voice belongs to or imitates a particular person
• Whether a decision is safe to make without further evidence

What is retained

The application sends the audio over HTTPS and keeps no site-side audio copy. Numeric results may be cached for up to 14 days against a file hash. There is no account history or public report link. Provider-side handling remains subject to the provider's terms.

Decide what evidence the stakes require

For personal curiosity, the report may answer enough. For moderation, contracting, copyright or discipline, obtain provenance such as drafts, session files and contributor records. Never make a consequential decision from this report alone. Review the known limitations or analyse another file.