Synced Lyrics

The goal is synced lyrics (LRC/Karaoke timing) for all of Lounès Matoub and many songs don’t even have written lyrics

Phase 1: Build the master discography

Start by creating a spreadsheet with:

Song	Album	Year	Audio source	Lyrics source	Status
Song name	Album	Release year	MP3/FLAC link	Existing text?	Missing / Transcribed / Timed

He recorded dozens of albums throughout his career, commonly cited as around 36 albums. (Wikipedia)

---

Phase 2: Collect every existing lyric

Before transcribing anything:

• Search lyric sites.

• Search fan forums.

• Search YouTube descriptions/comments.

• Search Facebook groups dedicated to Kabyle music.

• Search archived personal websites.

There are scattered lyrics available online, but coverage is incomplete. Sites such as LyricsTranslate and other lyric databases contain only a fraction of the catalog. (Lyrics Translate)

For each song:

• Save the lyrics exactly as found.

• Record the source.

• Mark confidence level.

---

Phase 3: Create missing lyrics from audio

This is the hard part.

For songs with no lyrics:

1. Separate vocals from instruments using an AI stem separator.

   • Demucs

   • UVR (Ultimate Vocal Remover)

2. Generate an initial transcript with a speech-to-text model.

   • Whisper large-v3

   • WhisperX

However, Kabyle ASR quality is currently poor. Community reports still mention that AI translations/transcriptions of Kabyle songs are often unreliable and require human correction. (Reddit)

So I would use AI only to create a rough draft.

3. Human correction.

Ideally:

• 2 native Taqbaylit speakers independently transcribe.

• Compare versions.

• Resolve disagreements.

This is where most of the quality comes from.

---

Phase 4: Standardize orthography

Kabyle lyrics online often use different conventions:

• Latin script with diacritics

• Simplified Latin script

• French-influenced spellings

Pick one standard.

I would recommend modern Kabyle Latin orthography:

• ɛ

• ɣ

• ṛ

• ḥ

• ṣ

• ṭ

• etc.

Then normalize all songs to that standard while preserving an archival copy of the original transcription.

---

Phase 5: Generate synced timings

Once plain lyrics exist:

Use:

• WhisperX forced alignment

• Aeneas

• Gentle

Workflow:

Audio
+
Corrected Lyrics
↓
Forced Alignment
↓
Word timestamps
↓
LRC generation

This gets you 80–95% of the way.

Then manually fix:

• Long instrumental sections

• Repeated refrains

• Spoken passages

• Audience chants

---

Phase 6: Quality-control pass

For each song:

• Listen while reading.

• Check every line start.

• Check every chorus repetition.

• Verify timestamps against waveform.

A trained editor can usually review a song in 5–15 minutes once the alignment exists.

---

Phase 7: Store everything in a structured archive

I would keep:

/Matoub
    /Album
        song.mp3
        song.txt
        song.lrc
        metadata.json

Example metadata:

{
  "title": "Ayen Ayen",
  "album": "...",
  "year": 1989,
  "lyrics_source": "manual",
  "confidence": 0.98
}

---

If I were doing the entire discography alone

My pipeline would be:

1. Download and catalog every recording.

2. Scrape all available lyrics.

3. Run Whisper on everything.

4. Correct transcripts manually.

5. Use WhisperX for alignment.

6. Export .lrc.

7. Review song-by-song.

The bottleneck is not synchronization—it’s producing accurate Kabyle lyrics for songs that have never been transcribed. Once the text exists, timing can be automated fairly effectively. The transcription stage will likely consume 80–90% of the total effort.