Song inspector

What each panel reads

Metadata is transcription, not inference. Fields that the page does not publish are marked “not published” rather than guessed — a missing play count means the page did not expose one, not that the song has none.

Style anatomy splits the style line on commas and slashes, then matches each part against small vocabularies for genre, instruments, vocal, mood, production and tempo. Parts that match nothing are kept as “unclassified” — that is useful signal, since a style line of mostly unclassified fragments is usually a line the model will read loosely. The panel also reports which of the six pillars had no match at all, which is the fastest way to see what a prompt forgot to say.

Lyric map reads [Section] tags and reports lines, words and the longest line per section. Long lines are where timing goes wrong; a chorus whose longest line is twice its average usually sings badly.

Audio profile only runs when you press the button. The file is downloaded, decoded with Web Audio in the tab, and measured: duration, sample rate, channel count, RMS level, true peak, crest factor, near-silence ratio and stereo correlation. If the fetch or the decode fails, the panel says so and the rest of the report stays intact.

Reading the audio numbers

• Crest factor — peak minus RMS. Small numbers mean a squashed, loud master; large numbers mean a dynamic one.
• Silence ratio — share of 20 ms windows below roughly −50 dBFS. A high value usually means dead air at the start or end rather than a quiet song.
• Stereo correlation — 1.0 is effectively mono, around 0 is wide, negative means the channels fight each other and the track will lose material when folded to mono.

Limits

This tool is read-only: it never publishes, edits or downloads on your behalf. It can only see what a public page exposes, so private songs return little. Analysis quality depends on the file the page links to — a low-bitrate preview measures like a low-bitrate preview.