Skip to content

Tags, variables and lyrics subtitles ​

Export Release writes the song's files; since 1.0.0 its tags and file names take variables, and it writes the lyrics as subtitles next to the song.

Tags ​

Choose tags in Export Release (the song templates do): artist, album, date, track, genre, comment, album artist, composer and lyrics. FLAC and MP3 carry them all (the lyrics as FLAC's LYRICS comment and MP3's USLT frame, which foobar2000, MusicBee, Poweramp and others show); WAV keeps only a few and no lyrics.

Every tag field takes typed text with variables, or a connected text: drag a link from any text output onto the field - a field output of the Song Brief, a writer's answer, a String node. The connected text wins over the typed one.

Variables ​

GroupVariables
song{title}, {lyrics}, {style}, {artwork_prompt}, {duration}, {bpm}, {score_key}
brief{description}, {template}, {mode}, {genre}, {mood}, {tempo}, {length}, {vocals}, {language}, {voice}, {theme}, {melody}, {lead_instrument}, {key}, {meter}, {arrangement}, {genre_closeness}, {lines_under_singing}, {creativity}; a cover's {harmony}, {song_flow_closeness}, {lyrics_closeness}
run{date}, {year}, {time}, {seed} (= {take_seed}), {draft_seed}, {arrangement_seed}
models{model}, {writer}
Plenio{plenio_version}
tags{artist}, {album}, {album_artist}, {composer}, {track}
  • {genre} - the value. The brief's fields come through Export Release's brief input (connected in the templates); a field left empty in the brief gives the template's value.
  • [ ... ] - the text inside appears only when every variable in it has a value: [Mood: {mood} · ] vanishes when the mood is empty. Brackets without a variable inside stay text ([Live]).
  • {{ and }} - literal braces.

Examples:

FieldTypedWritten
comment[Genre: {genre} · ][Mood: {mood} · ][Language: {language} · ]Powered by Plenio … (the templates')Genre: indie pop · Language: English · Powered by Plenio …
album{genre} Sessions {year}indie pop Sessions 2026
file name{artist} - {title}The Band - Harbour Lights.flac
file name{genre}/{date} {title}a folder per genre

A name that is no variable ({genr}) stops the run when you press Run - before anything renders - and the message lists the names. In connected text an unknown name stays as written (generated text may contain braces) and the export's summary notes it. The Tags & variables note beside Export Release in every template lists the variables, too.

Lyrics subtitles ​

lyrics subtitles (SRT) (on by default) writes three files next to the song:

  • <name> (lines).srt - one subtitle per lyrics line, shown from where its first word is sung until its last word ends;
  • <name> (notes).srt - the same lines painted on note by note: a new subtitle at every syllable (YuE2, on its note) or word (MiniMax), showing the line up to what is being sung - Hal, Halle, Hallelu, Hallelujah;
  • <name> (sung).srt - what the take really sings, as heard: a music model does not always sing its lyrics (it repeats a line, makes one up, leaves one out). The lines follow the lyrics wherever the take sings them - their line breaks, and their spelling where the same word was heard -, and the words the lyrics do not have make lines of their own. For a video of the song as it sounds, this is the file to use.

The release record keeps both - the lyrics' times, and what was sung with its times and its text with the lyrics' section tags - together with how far the take follows its lyrics: the share of their words that were heard, the words heard that they do not have, the lines not heard. The export's summary says it per take (86 % of the lyrics' words sung as written, 12 words the lyrics do not have) and names a section not sung as written; the Plenio Player follows either version.

What was heard stays as heard. The lyrics as sung are what the ASR heard - a word it misheard stands there as it was heard; nothing corrects them after the export. Misheard words are corrected only where the lyrics themselves come from an ASR: a cover's original lyrics, before the render (Fix Misheard Words).

How the times are measured. The lyrics ASR listens to the exported song. lyrics heard on (an option of Export Release) chooses the vocal stem (the default) or the song (quicker): Export Release then separates the take's vocals first (BS-RoFormer, the model of Stems and Sung Pitch - about 25 seconds for three to four minutes on a free GPU, longer while it is busy), so the music does not pull words into the instrumental parts. Measured on 20 songs against the vocals' own onsets and silence: words timed where nobody sings 2.5 % -> 0.3 %, words within 0.3 s of a sung onset 86.7 % -> 88.8 %. Without the separation model the song is heard, and the summary says so. Qwen3-ASR 1.7B hears the words and Qwen3-ForcedAligner 0.6B places every word on the audio (Apache-2.0, 4.1 + 1.8 GB, downloaded on first use; its results are cached). It hears the song in pieces of at most 20 seconds, cut at quiet moments, each with a second and a half of its neighbours (a word heard by both at a cut is kept once): over a whole song with its music the aligner loses its place - in 1.1.0, when it heard a song in one piece, lines could come up to a minute early. Measured on 15 YuE2 songs against faster-whisper, an independent timing: 94 % of the words within 3 seconds of Whisper's, 72 % of the word starts within 0.15 s of where the separated vocals set in. Its words are matched to the final lyrics - Chinese and Japanese character by character -; with a score (YuE2), the score is fitted to the heard words and every syllable takes its note's time - the heard time wins where the two disagree. A YuE2 render is shifted against its score (1.7 s in the median song), so the score's own times are not used. For a language the aligner does not know (it times English, German, French, Spanish, Italian, Portuguese, Russian, Chinese, Cantonese, Japanese and Korean) faster-whisper large-v3 (MIT, 3.1 GB) takes over and the export's summary says so; its word times start early and are corrected - measured on 48 YuE2 songs against the aligner: half the words within 0.16 s, 82 % within 0.5 s, 92 % within 1 s.

  • Without lyrics (Enhance & Master): the subtitles are the words as heard, a line per phrase; the sentences an ASR is known to invent into music are left out, and without singing there are none.
  • Instrumentals: no subtitles.
  • A line nobody sang (the render left it out) is placed between its neighbours when there is room, otherwise left out; the summary names it.
  • Offline without the model, or when the ASR fails: the export goes on without subtitles and says why.

The files use no styling tags: SRT markup is read differently by players, YouTube and video editors, so the files work everywhere. Into DaVinci Resolve: File > Import > Subtitle…, then drag the subtitle track onto the timeline at the song's start; YouTube: Subtitles > Upload file > With timing.