← Back to Vocal Check

Real audio. Recorded results.

The built-in demo uses four licensed MP3 excerpts and the same detector used for your own queue. These are actual browser measurements, saved on 2026-09-17.

RecordingVoice scoreDecisionStrongest class
Let's Go Fishin'19.74%reviewSinging
Horses0.35%lowSpeech
The Ashiel Mystery · narration81.67%voiceSpeech
The candle0.38%lowSpeech

The song is a borderline result sent for review. Scores are uncalibrated model evidence; a low score does not certify an instrumental. Slight decoder/backend differences can move a score near the threshold between review and voice-likely.

The full Musefi batch

All 217 URLs were attempted. 203 recordings were analyzed; 14 audio URLs remained unreachable after a retry. Failed downloads have no confidence score. Each successful recording was sampled for up to 30 seconds. The first pass and retry took 512 seconds of batch execution.

This is a dated snapshot, exported 2026-09-17T12:26:17.364Z. Scores are screening evidence, not listener-confirmed vocals or measured detection accuracy.

Full JSON audit · Full CSV audit · Open all batch results instantly

What the demo proves

The controls exercise a sung recording, spoken narration, and two Musefi excerpts. This is a functional smoke test, not a labelled catalogue benchmark. The two Musefi excerpts have not been independently labelled by a listener. The requester’s four listener-confirmed problem tracks were not supplied.

Website checks

Reproduce it

Open the four-track demo and click Run demo scan. The first run downloads the model; processing time depends on your device. Keep the tab open. Use the included command-line scanner for larger queues.

Download detailed measured results · Audio sources and licences

AST AudioSet model · ONNX q8 · model revision 249a1fbf0286b40e7f1ed687a8ae396997bf7dc6 · ROKKO Works