Add lyrics pipeline, playlist views, admin toggles, and player polish

Lyrics pipeline (Whisper + Demucs + description alignment):
- New GenerateLyricsJob runs WhisperX with VAD filtering and forced word
  alignment, writes per-track JSON to NAS.
- New DecorateLyricsJob calls the active LLM provider to bake one to
  several emojis into each line (heavy decoration prompt).
- LyricsDescriptionParser strips heading content, section markers, and
  emoji-decoration from a song's description while preserving every
  language verbatim.
- correct_whisper_with_description aligner: strong-match anchors only,
  vocal-region-aware gap-fill so missing verses land on actual singing.
- Owner UI for generate/regenerate/edit/delete in the player gear.

Admin pages:
- /admin/lyrics    toggles for VAD, vocal gap-fill, Demucs, master
- /admin/gpu       extracted GPU section, encoder picker, FFmpeg path
- /admin/backup    extracted users-and-settings export/import
- /admin/settings  now AI/LLM only with provider list and Test button
- /admin/nas-storage hosts NAS settings, repair, disable flow, browser
- Shared partials/settings-styles for a uniform look across pages.

Playlist view tracking:
- Migration adds playlists.view_count and playlist_views dedup table.
- Playlist::bumpViewIfNew increments per device with a one-hour window.
- Tracked from /playlists/{id}, /playlists/share/{token}, /ps/{token},
  and /videos/{id}?playlist={token}.  Dispatched after-response so it
  never blocks the page render.
- Loading a playlist on the video page now runs one query instead of
  the four the old getNextVideo/getPreviousVideo path triggered.
- View counts shown on every playlist card and the playlist hero.

Player polish:
- Floating mini-player is draggable, persists its position in
  localStorage, clamps to viewport on resize.
- Mini disabled entirely on mobile (less than 768px).
- New gear-menu Mini Player toggle (persists in localStorage) lets the
  user disable both scroll-activation and SPA-nav-activation.
- Close button keeps media playing when used on the player's own page.
- SPA navigator now swaps a #page-scripts container so per-page JS
  (channel tabs, etc.) gets re-executed after content swaps.

Storage layout:
- Runtime data moved from /storage/* to /data/* and gitignored.
- /ml/venv, /ml/cache, /ml/__pycache__ excluded.

Co-Authored-By: Claude Opus 4.7 <[email protected]>
This commit is contained in:
ghassanandClaude Opus 4.7 committed 2026-05-31 22:01:47 +03:00
1 parent 73527f3781
commit f98e5415a3
60 files changed
+7920 -2675

No files matched your search

+144
View File
@@ -0,0 +1,144 @@
<?php
namespace App\Support;
/**
* Extract clean lyric lines from a video's free-text description.
*
* Many users paste the song's lyrics into the description with section markers,
* emoji decorations, instrument tags, etc. When that's present we want to USE
* those exact lines (and let the pipeline only do the *sync*, not the
* transcription), because they're far more accurate than anything Whisper can
* derive from sung audio.
*
* Returns an empty array when no usable lyric block is found.
*/
class LyricsDescriptionParser
{
/** Heuristic threshold: descriptions with fewer cleaned lines aren't worth aligning. */
private const MIN_LYRIC_LINES = 4;
public static function extract(?string $desc): array
{
if (! $desc) return [];
// HTML descriptions use <br> for line breaks — convert those (and other
// block-ending tags) into real newlines BEFORE stripping tags, otherwise
// the entire body collapses into one long run-on line.
// Heading tags (<h1>…<h6>) carry the song title — drop their content
// entirely so the title never leaks into the lyric list.
$text = preg_replace('/<\s*h[1-6][^>]*>.*?<\s*\/\s*h[1-6]\s*>/isu', "\n", $desc);
$text = preg_replace('/<\s*br\s*\/?>/i', "\n", $text);
$text = preg_replace('/<\s*\/\s*(p|div|li|tr|blockquote)\s*>/i', "\n", $text);
$text = strip_tags($text);
// Decode HTML entities (&nbsp;, &amp;, etc.) so the comparison later isn't fooled.
$text = html_entity_decode($text, ENT_QUOTES | ENT_HTML5, 'UTF-8');
$text = preg_replace('/\r\n|\r/u', "\n", $text);
// First pass: clean each line and flag section headers (Verse / Ritornello
// / Bridge / etc.) so they can be dropped — those aren't sung.
$cleaned = [];
foreach (explode("\n", $text) as $line) {
$line = self::cleanLine($line);
if ($line === null) continue;
$cleaned[] = ['text' => $line, 'header' => self::isSectionHeader($line)];
}
// Title detection: if the first non-header line is immediately followed by
// a section header (e.g. "Figlio Mio — Viaggio di Vita" then "Verso 1"),
// that first line is the song title — drop it too.
$firstIdx = null;
foreach ($cleaned as $i => $c) {
if (! $c['header']) { $firstIdx = $i; break; }
}
$dropTitle = false;
if ($firstIdx !== null) {
for ($j = $firstIdx + 1; $j < count($cleaned); $j++) {
if ($cleaned[$j]['header']) { $dropTitle = true; break; }
break; // first thing after is a real lyric line → not a title block
}
}
$out = [];
foreach ($cleaned as $i => $c) {
if ($c['header']) continue;
if ($dropTitle && $i === $firstIdx) continue;
$out[] = $c['text'];
}
// Avoid mistakenly aligning non-lyric descriptions (a credit line, a URL,
// etc.). Require at least a handful of plausible lyric lines.
if (count($out) < self::MIN_LYRIC_LINES) return [];
return $out;
}
/**
* True when a line is a section marker (Verse / Chorus / Bridge / Outro /
* their many translations) rather than a sung lyric. Matches the WHOLE line
* so a real lyric containing one of these words isn't mistakenly dropped.
*/
private static function isSectionHeader(string $line): bool
{
$t = mb_strtolower(trim($line));
if ($t === '') return false;
$roots = [
'intro', 'outro', 'interlude', 'instrumental',
'verse', 'verso', 'verset', 'couplet', 'estrofa', 'strofa',
'chorus', 'ritornello', 'refrain', 'refrão', 'refrao', 'coro', 'estribillo',
'pre[\s\-]?chorus', 'pre[\s\-]?ritornello', 'pre[\s\-]?coro',
'pre[\s\-]?refrain', 'pre[\s\-]?refrão', 'pré[\s\-]?refrain',
'bridge', 'ponte', 'puente', 'pont', 'brücke', 'brucke',
'hook', 'drop', 'breakdown', 'tag', 'vamp', 'coda', 'reprise',
// CJK / Thai / Arabic / Korean
'サビ', 'コーラス', 'バース', 'ブリッジ', 'イントロ', 'アウトロ', 'フック', '間奏',
'ท่อน', 'คอรัส', 'ฮุก', 'บริดจ์', 'อินโทร', 'เอาท์โทร',
'前奏', '副歌', '桥段', '主歌', '尾奏',
'كورس', 'بريدج', 'كوبليه',
'후렴', '브릿지', '인트로', '아웃트로', '훅',
];
$rootRe = implode('|', $roots);
// Optional trailing number, "final/finale/reprise", roman numerals.
$pattern = '/^(?:' . $rootRe . ')[\s\d:\-—\.]*(?:final|finale|reprise|ii|iii|iv|v|vi|2|3|4|5)?\s*$/iu';
return (bool) preg_match($pattern, $t);
}
/** Returns the cleaned line, or null if it should be discarded. */
private static function cleanLine(string $line): ?string
{
$line = trim($line);
if ($line === '') return null;
// Strip markdown emphasis (* _ ~) and leading list bullets / quote markers.
$line = preg_replace('/^[\s>\-\*•♪♫·]+/u', '', $line);
$line = preg_replace('/[\*_~`]+/u', '', $line);
// Drop instrument / section annotations inside Japanese-style brackets:
// 【 箏・尺八・篠笛・優しい歌声 】 — these aren't lyrics.
$line = preg_replace('/【[^】]*】/u', '', $line);
$line = preg_replace('/〔[^〕]*〕/u', '', $line);
$line = preg_replace('/\[\[[^\]]*\]\]/u', '', $line);
// Strip emoji / pictographic symbols and the invisible glue that often
// sticks to them (variation selectors, ZWJ) so nothing leaves behind a
// bare diacritic when the visible emoji is removed.
$line = preg_replace(
'/[\x{1F000}-\x{1FFFF}\x{2600}-\x{27BF}\x{2B00}-\x{2BFF}\x{0F3A}-\x{0F3D}\x{FE00}-\x{FE0F}\x{200B}-\x{200F}\x{2060}]/u',
'', $line
);
// Collapse internal whitespace.
$line = preg_replace('/\s+/u', ' ', $line);
$line = trim($line);
if ($line === '') return null;
// Must contain at least one letter (Unicode), and at least 3 characters
// after stripping — discards "🌸 平穏 🌸" (header) and "──" separators.
if (! preg_match('/\p{L}/u', $line)) return null;
if (mb_strlen($line) < 3) return null;
return $line;
}
}